Software Agent Oriented Programming

CHAPTER 13

An Overview of
Agent-Oriented Programming

Yoav Shoham

have been working in areas related to software agents for a number of years

1 now, together with many students and other colleagues. Recently, terms
such as "(intelligent) (software) agents," "knowbots," and "softbots" have
become quite popular. The work taking place under this umbrella is di verse,
varying in content, style, and quality sufficiently to render terms such as "soft-
ware agent" meaningless in general. I have spent a fair amount of time in the
past two years trying to understand various agent-related work in industry and
academia. However, in this chapter I will not attempt to put any order into this
area, nor position our own work at Stanford within it. This is the topic of an-
other paper currently in the works. The discussion here will be confined to re-
viewing our own work on multi-agent systems in general and agent-oriented
programming in particular.

Agent-Oriented Programming: Software with Mental State

In 1989 I coined the term agent-oriented programming (AOP) to describe a new
programming paradigm, one based on cognitive and societal view of computa- I
tion. Although new, the proposal was inspired by extensive previous research in I
,1
Artificial Intelligence (Al), distributed computing, and other neighboring disci-
plines. This chapter will summarize some of the major ideas from previous re-
search. A more detailed discussion of AOP appears in Shoham (1993).
,,11
II~
What Is an Agent?
I~il
Most often, when people in Al use the term "agent," they refer to an entity that

AGENT-ORIENTED PROGRAMMING 273
272 SHOHAM

functions continuously and autonomously in an environment in which other
1 ties or qualities isomorphic to them. Theories of belief, knowledge and wanting 111't
can be constructed for machines in a simpler setting than for humans, and later ap- IJ~
processes take place and other agents existo This is perhaps the only property
plied tO humans. Ascription of mental qualities is most straightforward for machines '''1Ii
that is assumed uniformly by those in Al who use the termo The sense of "au- ,
"Ii'
of known structUre such as thermostats and computer operating systems, but is
tonomy" is not precise, but the term is taken to mean that the agents' activities
most useful when applied to entities whose structure is very incompletely known. I~i
do not require constant human guidance or intervention. Often certain further ,.UIII1
In Shoham (1989), I illustrate the point through the light-switch example. Ir ,!~r~~
assumptions are made about the environment, for example that it is physical 'Ii''''!
and partially unpredictable. lo fact, agents are sometimes taken to be robo tic is perfectly coherent to treat a light switch as a (very cooperative) agent with the
agents, in which case other issues such as sensory input, motor control, and time capability of transmitting current at will, who iovariably transmits current
pressure are mentiooed. when it believes that we want it traosmitted and not otherwise; flicking the
Fioally, agents are often taken to be "high-level." Although this sense is quite switch is simply our way of communicating our desires. However, while this is
vague, many take some version of it to distioguish agents from other software or a coherent view, it does not buy us anything, sioce we esseotially understand the
hardware components. The high level is manifested io symbolic representation mechanism sufficiently to have a simpler, mechanistic description of its behav-
and/or some cognitive-like function: agents may be "informable" (Geoesereth ior. In cootrast, we do not have equally good knowledge of the operation of
1989), may contain symbolic plans in addition to stimulus-response rules (Tor- complex systems such robots, people, and, arguably, operating systems. lo these
rance 1991;Hayes-Roth et al. 1989; MitcheIl1990), and may eveo possess oatural- cases it is often most convenient to employ mental terminology; the application
language capabilities. This sense is not assumed uniformly in Al, aod in fact a of the concept of "knowledge" to distributed computation, discussed below, is
certain counter-ideology deliberately denies the centrality or even existence of an example of this convenience.1
high-Ievel representation in agents (Agre and Chapman 1987; Brooks 1986).
Clearly, the notion of agenthood in Al is anything but crispo I should there- Agent- Versus Object-Oriented Programming
fore make it cIear what 1 mean by the term "agent," which is precisely this: An I mentioned previously that the ascription of mental constructs must be coher-
agent is an entity whose state is viewed as consisting of mental components such as ent and usefuI. The application of the logic of knowledge in distributed compu- 11

beliefs, capabilities, choices, and commitments. These components are defined io a 11
tation, given there as an example, used the mental construct "knowledge" in a
precise fashion and stand in rough correspondence to their commonsense coun-
particular way: it mapped it onto an existing computational framework (a dis-
terparts. In this view, therefore, agenthood is in the mind of the programmer: tributed network of processors) and used it to reason about the system. The use
What makes any hardware or software component ao agent is precisely the fact we will make of mental constructs is different: rather than use them for mere
that one has chosen to analyze and control it in these mental terms. 11_11

analysis, we will employ them to design the computational system. The various
The question of what an agent is is oow replaced by the questioo of what en- l."
mental categories will appear in the programming language itself, and the se-
tities can be viewed as having mental state. The answer is that anything can be mantics of the programming language will be related to the semantics of the
'111

so described, although it is not always advantageous to do so. This view is not mental constructS. This is similar in spirit to a development within the dis- III~'

original to me. For example, in Dennett (1987) and other publicatioos, Dennett tributed computation community, where a proposal has been made to include
011

proposes the "intentional stance," from which systems are ascribed mental qual-
tests for epistemic properties in the protocols themselves (Halpern and Zuck 111
ities such as intentions and free will. The issue, according to Dennett, is oot
1987); however, up till now there has been no follow up on the proposal. 111

whether a system really is intentional, but whether we can coherently view it as
I have proposed a computational framework called agent-oriented program- "1,

such. Similar sentiments are expressed by McCarthy (1979), who also distin- 11
ming (AOP). The name is not accidental, since from the engineering point of
guishes between the legitimacy of ascribing mental qualities to machines and its view AOP can be viewed as a specialization of the object-oriented programming
usefulness: 11:

(OOP) paradigm. I mean the latter in the spirit of Hewitt's original Actors for-
To ascribe certain beliefs, free will, intentions, consciousness, abilities, or wants to a
malism (Hewitt 1977), rather than in some of the senses in which it used today.
machine or computer program is legitimate when such an ascription expresses the
Intuitively, whereas OOP pro poses viewing a computational system as made
same information about the machine that it expresses about a persono It is useful
up of modules that are able to communicate with one another and that have in-
when the ascription helps us understand the structure of the machine, its past or
dividual ways of handling incoming messages, AOP specializes the framework 1'1

future behavior, or how to repair or improve it. It is perhaps never logically re-
quired even for humans, but expressing reasonably briefly what is actualIy known by fixing the state (now called mental state) of the modules (now called agents) to
about the state of the machine in a particular situation may require mental quali- consist of precisely defined components called beliefs (including beliefs about

""""

274 SHOHAM
! AGENT-ORIENTED PROGRAMMING 275

OOP AOP the resulting systems wilI be "if agent A believes that agent B has chosen to do
Basic unit object agent something harmful to agent A, then A will request that B change its choice." In
unconstrained
addition, temporal information is included to anchor belief, choices, and so on
Parameters defining beliefs, commitments,
state of basic unit in particular points in time. 11
capabilities, choices...
Here again we may benefit from some ideas in philosophy and linguisrics. As ~~~
Process of computation message passing and in the case of knowledge, there exists work in exact philosophy on logics for
message passing and
response methods response methods choice and ability. Although they have not yet had an effect in Al comparable to
Types of message unconstrained
that oflogics ofknowledge and belief, they may in the future.
inform, request, offer, Intentional terms such as knowledge and belief are used in a curious sense in
promise, decline...
the formal Al community. On the one hand, the definitions come nowhere cIose
Constraints on methods none to capturing the fulI linguistic meanings. On the other hand, the intuitions
honesty, consistency...
about these formal notions do indeed derive from the everyday, common sense
meaning of the words. What is curious is that, despire the disparity, the every-
Figure 1. OOP versus AOP. day intuition has proven a good guide to employing the formal notions in some
circumscribed applications. AOP aims ro strike a similar balance between com-
putational utility and common sense.
the world, about rhemselves, and abour one anorher), capabiliries, choices, and
possibly orher similar norions. A compurarion consists of rhese agents' inform-
ing, requesting, offering, accepting, rejecting, competing, and assisting one an- Two Scenarios
other. This idea is borrowed directly from the speech act literature (Grice 1989;
Searle 1969; Austin 1962).
Below are rwo scenarios. The first is fairly complex and serves to iIIustrate the
Speech act theory categorizes speech, distinguishing between informing, re- rype of future applications envisioned. The second is a toy example and serves
questing, offering and so on; each such type of communicative act involves dif- rhree purposes: it iIIustrates a number of AOP ideas more crisply; it is imple-
ferent presuppositions and has different effects. Speech-act theory has been ap- mentable in the simple AGENT-O language described later in the chapter; and
plied in Al, in naturallanguage research as welI as in plan recognition. To my it ilIustrates the fact that agents need not be robotic agents.
knowledge, AOP and McCarthy's Elephant2000 language are the first attempts Manufacturing AutomatÍon. Alfred and Brenda work at a car-manufacturing
to base a programming language in part on speech acts. Figure 1 summarizes plant. Alfred handles regular-order cars, and Brenda handles special-order ones.
the relation between AOP and 00p'2
The plant has a welding robor, Calvin. The plant is controlIed by a coordinating
program, Dashiel. The folIowing scenario develops, involving communication be- I1
On the Responsible Use ofPseudo-Menral Terminology tween Alfred, Brenda, Calvin and Dashiel. Ir contains communication acts such
The previous discussion referred to mentalistic notions such as belief and com- as informing, requesting, committing, permitting, and commanding and requires I
mitment. In order to understand the sense in which I intend these, consider the agents to reason about the beliefs, capabilities, and commitments of other agents.
use of logics of knowledge and belief in Al and distributed computation. These 8:00: Alfred requests that Calvin promise to weld ten bodies for him that
I
logics, which were imported directly from anal ytic philosophy firsr to Al and day; Calvin agrees to do so.
then to other areas of computer science, describe the behavior of machines in 8:30: Alfred requests thar Calvin accept the first body, Calvin agrees, and
terms of notions such as knowledge and belief. In computer science these men- the first body arrives. Calvin starts welding it and promises Alfred to noti-
talistic-sounding notions are actualIy given precise computational meanings and fy him when it is ready for the next body.
are used not only to prove properties of distributed systems, but to program 8:45: Brenda requests that Calvin work on a special-order car which is
them as well. A typical rule in such a knowledge-based systems is "if processor needed urgently. Calvin responds that it cannot righr then but that it will ~
A does not know that processor B has received its message, then processor A when ir finishes the current job, at approximately 9:00.
will not send the next message." AOP augments these logics with formal no- 9:05: Calvin completes welding Alfred's first car, ships it out, and offers to
tions of choices, capabilities, commitments, and possibly others. A typical rule in weld Brenda's caroBrenda ships it the car, and Calvin starts welding.

276 SHOHAM

9:15: Alfred inquires why Calvin is not yet ready for his (Alfred's) next caro
Calvin explains why and also that it (Calvin) expects to be ready by about
10:00.
l P fa C: Please get permission to do so.
C fa S: I request permission for the following double booking.
"

!I
I
S fa C: Permission denied.
9:55: Calvin completes welding Brenda's car and ships it out. Brenda re- C fa P: Sorry, I cannot get approval.
quests that it reaccept it and do some painting, but Calvin refuses, explain-
April18, at the airport
ing that it does not know how to paint. Calvin then offers to weld another
P fa C: My name is P; I have a reservation for flight 293.
car for Alfred and proceeds to weld Alfred's cars for a while.
C fa P: Here is your boarding pass.
12:15: Brenda requests that Calvin commit to welding four more special-
order cars that day. Calvin replies that it cannot, since that conflicts with its Overview of the AOP Framework
commitment to Alfred, who still has six unwelded cars. Brenda requests
Alfred to release Calvin from its commitment to Alfred. Alfred refuses.
Brenda requests that Dashiel (remember Dashiel?) order Calvin to accept
her important request and revoke its commitment to Alfred. Dashiel or-
.
A complete AOP system will include three primary components:
A restricted formallanguage with cIear syntax and semantics for describ-
ing mental sta te. The mental state will be defined uniquely by several
ders Calvin to weld two of Brenda's cars, and then as many of Alfred's as modalities, such as belief and commitment.
time allows.
3:30: Calvin senses it is low on lubricant, and, without interrupting its
. An interpreted programming language in which to program agents, with
primitive commands such as REQUEST and INFORM. The semantics of
work, asks Alfred for when the next lubrication is scheduled. Alfred re- the programming language will depend in part on the semantics of mental
sponds that he doesn't know, but that Dashiel should know. Calvin asks state.
Dashiel, who informs it that the next lubrication is scheduled for 4:00.
Since that is sufficiently soon, Calvin continues to weld.
. An "agentifier," converting neutral devices into programmable agents.
In the remainder of this document I will start with a short discussion of men-
Airline Reservation. This example is a minor modification of one due to John tal state. I will then present a general family of agent interpreters, a simple rep-
McCarthy (1990), who uses it to illustrate his Elephant2000 programming lan- resentative of which has already been implemented as AGENT-O. Related work
guage. The example involves three agents: a passenger P, an airline cIerk C, and is described in the fifth section. I will end with a summary of recent research re-
C's supervisor S. P and S are human, and C is a program (or is it the other way sults related to AOP.
around? Or perhaps all three are programs?). Note that by confirming a reser-
vation, the airline enters ioto a commitment to issue a boarding pass to the pas-
senger at the appropriate time. Again, the following exchange involves request- On the Mental State of Agents
ing (a query is a request to inform), informing, permitting, and denying, as well
as mental components such as knowledge and commitment. The first step in the enterprise is to define agents; that is, to define the various
March componeots of mental state and the interactions between them. There is not a
P fa C: Please inform me what flights you have from San Francisco to New unique "corred' definition, and different applications can be expected to call
York on April18. for specific mental properties.3 In this section 1 summarize what could be
viewed as a bare-bones theory of mental state, a kernel that will in the future be
C fa P: Flight 354 departs at 08:30, flight 293 departs at 10:00,flight 441 de-
modified and augmented.
parts at noon.
P fa C: Please book me on 354.
Components of a Language for Mental State
C fa P: That is sold out.
P fa C: Please book me on 293. In related past research by others in Al, three modalities were explored: belief,
desire, and intention (giving rise to the pun on BDI agent architectures). Other
C fa P: That is confirmed; your reservation number is 112358. similar notions, such as goals and plans, were also pressed into service. These
P fa C: Please book me also on 441. are cIearly important notions; they are also complex ones, however, and not nec-
C fa P: That conflicts with 293; I am not allowed to double book a passenger. essarily the most primitive ones.4

278 SHOHAM

By way of motivation, here is an inliCo __8Jla1view of the world which underlies These properties are quite weak, but they are sufficient to justify the termi-
the selection. At any point in time, the:::: FLJlUre is determined by two factors: the nology, and necessary for the design of the interpreter. The weakness of the as-
past history, and the current actions ~f =a..gents. For example, past history alone sumptions ensures that the interpreters apply to a wide variety of applications.
does not (in this view) determine whetl:~c:: rI raise my arm; that is determined by StilI, even these assumptions wilI be inappropriate for some purposes, in which
whether in fact I take the appropriate auc=_ion. The actions of an agent are deter- case a new type of interpreter wilI be required.
mined by its decisions, or choices.s In ot::h oe:.words, some facts are true for natural Internal consistency. I assume that both the beliefs and the obligations are in-
reasons, and other facts are true becalLse:: ~gents decided to make them so. Oeci- ternalIy consistent.
sions are logically constrained, thoug h J;;:1otdetermined, by the agent's beliefs; Good faith. I further assume that agents commit only to what they believe
these beliefs refer to the state ofthe wc:>r-.í:l.d(in the past, present or future), to the themselves capable of, and only if they realIy mean it.
mental state of other agents, and to the:: '<::8apabilities of this and other agents. For Introspection. Although in general I do not assume that agents have total in-
example, given that the robot believes tl atit is incapable of passing through the trospective capabilities, I do assume that they are aware of their obligations. On
the other hand, I do not assume that agents are necessarily aware of commit-
narrow doorway, it will not decide t, ~ go through it. Oecisions are also con-
ments made to them.
strained by prior decisions; the robot C~LIIIot decide to be in Room 5 in five min-
utes ifit has already decided to be in R QaCJ8113 at that time. Persistenceof mental state. I have only placed restrictions on mental attitudes
In the first instantiation of AOP, a 1 anguage calIed AGENT-O (Torrance at a single instant of time. I conclude this section by discussing how mental
1991), we too started with quite basic b L.L ilding blocks, in fact much more basic states change or persist over time. Unlike with the previously discussed proper-
than those mentioned so far. We incor~~rated two modalities in the mental state ties, precise constraints cannot currently be specified, but only informal guide-
lines.
of agents: belief and obligation (or co fY18J,..,Útment). We also defined decision (or
Consider, for example, belief. The previously discussed restrictions allow
choice) as an obligation to oneself. FinaalL.y-,lve included a third category which is
not a mental construct per se: capability . agents which at one time believe nothing at alI, shortly afterwards to have a be-
lief about every sentence, and then again to become quite agnostic. Common
By restricting the components of m-e~t::al state to these modalities we in some
informal sense excluded representation... (]ID motivation. Indeed, we did not assume
f sense suggests that beliefs tend to be more stable than that, and it would indeed
be difficult to rely on the behavior of agents with such volatile beliefs. I wilI now
that agents are "rational" beyond assulIl.:i rlg that their beliefs, obligations and ca-
pabilities are internalIy and mutually c::o ~sistent. This assumption stands in con- place a strong condition on belief: I wilI assume that agents have perfect memo-
trast to the other work mentioned a~c:.""'e, which makes further assumptions ry of and faith in their beliefs, and only let go of a belief if they learn a contra-
dictory fact. Beliefs therefore persist by default. Furthermore, I wilI assume that
about agents acting in their own best i nt;'=:rests, and so on. Such stronger notions the absence of belief also persists by default, although in a slightly different
of rationality are obviously important, a J. <J[ am convinced that in the future we
sense: if an agent does not believe a fact at a certain time (as opposed to believ-
will wish to add them. In fact, in her .o.::s::s:rtation, Thomas introduced an AOP
ing the negation of the fact), then the only reason he wilI come to believe it is if
language that includes the notions of ir» t= rlding and planning (Thomas 1993).
he learns it.
How to formally capture these two kinds of default persistence is another
Properties of the Various Components
story, which touches on issues that are painfulIy familiar to researchers in non-
I have so far not placed any constrain t> eun the various modalities defined, and monotonic temporal reasoning and belief revision. In fact, a cIose look at the
therefore ha ve not guaranteed that t81=}oI in any way resemble their common logical details of belief (or knowledge) persistence reveals several very subtle
sense counterparts. We will now place s c:h constraints. Just as there is no objec- phenomena, which have so far not been addressed in the literature.
tively "right" collection of mental catega::D r-ies, there is no "right" list of properties In addition, obligations should persist--otherwise they wouldn't be obliga-
for any particular mental category. I r- .<lavealready stated that the correspon- tions. As in the case of belief, however, the persistence is not absolute. Although
dence between the formal definition aIlc::8 common sense wilI always be only ap- by default obligations persist, there are conditions under which obligations are
proximate and that I would like to str.lc~ abalance between common sense and revoked.
utility. Indeed, I expect different appli.occa t:ions of AOP to calI for different prop- These conditions presumably include explicit release of the agent by the party
erties of belief, commitment, and capa o. :J:iiry.In this section [ wilI brieAy and in- to which it is obligated, or alternatively a realization on the part of the agent
formalIy define a number of properties ]!::assume about the modalities. Formal that it is no longer able to fulfilI the obligation. (In their discussion of the persis-
definitions of these properties may be FO~.;I1d in Shoham (1993). tence of commitment, Cohen and Levesque [1990] actualIy propose a more elab-

,
AGENT-ORIENTED PROGRAMMIN(; 281
280 SHOHAM

orate second condition, one that requires common knowledge by the committer
and committee of the impossibility; however, further discussion of their position Initialize mental state
and capabilities
and arguments against it would be too long a detour.)
Define rules for making
Since decision is defined in terms of obligation, it inherits the default persis- new commitments
tence. Notice, however, an interesting point about the persistence of decision:
while an agent cannot unilaterally revoke obligations it has towards others, it
can cancel obligations held towards it-ineluding obligations it holds towards
itself, namely decisions. An agent is therefore free to modify an existing deci-
, ~-
sion, but unless he explicitly does so the decision will stand. --~ '"
UpJate
Finally, capabilities too tend not to Auctuate wildly. In fact, in this document Representation of
I assume that capabilities are fixed: What an agent can do at one time it can do dock - - mental
'-1 mentat state
and capability
sta te
at any other time. However, I will allow to condition a capability of an action on 1
certain conditions that hold at the time of action.
I A
The Con textual Nature of Modal Statements

I have throughout the discussion talked of "unequivocal" statements regarding
beliefs, obligations, and capabilities. Common sense, however, suggests that
+ ,I

Execute commitments
fOTcurrent time
, I
I
-.
each of these modalities is context sensitive: I can print the document right now, ()utgoing messages
but only in the context of the network being up; I am obligated to you to finish
the work by tomorrow, but if my child has just been rushed to hospital then all
control -...
bets are off (even though I am still capable of finishing the work). Indeed, Mc-
Carthy has argued that all statements, not only modal ones, should be viewed in
contexto Although I agree in principie and discuss it further in Shoham (1991),
in this artiele I will ignore the issue of context sensitivity.
Figure 2. A flow diagram ola generic agent interpreter.

A Generic Agent Interpreter

In the previous section I discussed the first component of the AOP framework, 2. Execute the commitments for the current time, possibly resulting in fur-
namely the definition of agents. I now turn to the programming of agents and ther belief change. Actions to which agents are committed inelude com-
will outline a generic agent interpreter. municative ones such as informing and requesting.
The behavior of agents is governed by programs; each agent is controlled by
The process is illustrated in figure 2; the dashed arrows represent Aow of
its own private programo Agent programs themselves are not logical entities,
data, while the solid arrows show temporal sequencing.
but their control and data structures refer to the mental state of the agent using
the logicallanguage.6
Assumptions about Message Passing
The Basic Loop Agent programs will inelude, among other things, communication commands.
In order that those be executable, I will assume that the platform is capable of
The behavior of agents is, in principie, quite simple. Each agent iterates the fol-
lowing two steps at regular intervals: passing messages to other agents addressable by name, whether those reside in
the same machine or in others. The programming language will define the form
1. Read the current messages and update your mental state (ineluding your
beliefs and commitments); of these messages, and the interpreter will determine when messages are sent.

282 SHOHAM

Related W ork ''Ii
Assumption about the Clock
I~~
Central to the operation of the interpreter is the existence of a dock; agents are ~~
So far, I have not discussed related work in any depth. The body of related t~ii
inherently "real time" (to use another overloaded term). The main role of the work is in fact so rich that in this section I will mention only the most dosely re-
dock is to initiate iterations of the two-step loop at regular intervals (e.g., every lated work, and briefIy at that. I will omit further discussion of past work on '1'
10 milliseconds, every hour). The length of these intervals is determined by the t~
logics of knowledge and belief, which the logic of mental state extends, since I ~~
settable variable "time grain." already did that in the introduction. For the same reason, I will not discuss ob-
I do not discuss the implementation of such a dock, which will vary among
ject-oriented programming and Hewitt's work. The following is ordered in
platforms, and simply assume that it exists. I also assume a variable "now," what I see as decreasing relevance to, and overlap with, AOP. The order (or, for
whose value is set by the dock to the current time in the format defined in the that matter, indusion in the list) refIects no other ranking, nor is it implied that
programming language (e.g., an integer, date:hour:minute). researchers high up on the list would necessarily endorse any part of AOP.
In previous work, I have made the very strong assumption that a single itera-
tion through the loop lasts less than the time grain; in future versions of the lan-
McCarthy's (1990) work on Elephant2000
guage I will relax this assumption and correspondingly will complicate the de-
tails of the loop itself. . This language under development is also based on speech acts, and the airline-
Of course, the fact that agents use the same temporallanguage does not en- reservation scenario I have discussed is due to McCarthy. One issue explored in
sure that their docks are synchronized. If all are agents are running on the same connection with Elephant2000 is the distinction between illocutionary and per- 111

machine; there will be no problem, but otherwise the possibility of dock drift locutionary specifications, which 1 have not addressed. In contrast to AOP, Ele- .111

exists. Although synchronization does not impact the design and programming phant2000 currently contains no explicit representation of state, mental or oth-
of single agents, it is crucial for ensuring that a society of agents is able to func- erwise. Conditional statements therefore refer to the history of past
tion usefully. Fortunately, there exist synchronization protocols which ensure communication rather than to the current mental state.
1111

limited drift among docks (for an overview, see Schneider [1987]), and we ex-
11
pect to use these in our applications. Distributed Al
'1111

There is related work within Distributed Al community (cf. MCC [1990]). Al- .111

AGENT-O: A Simple Language and its Interpreter though AOP is, to my knowledge, unique in its definition of mental state and
the resulting programming language, others too have made the connection be- '1

tween object-oriented programming and agenthood (Ferber and Carie 1990;
Agent interpreters may vary along many dimensions and in general pose many
Hewitt 1990). 111

challenging problems. We have implemented a particular programming lan-
guage called AGENT-O, whose interpreter is an extremely simple instance of
The Intelligent Communicating Agents Project (1987-1988)
the generic agent interpreter. In fact, the simplifications embodied in AGENT-
"
Oare so extreme that it may be tempting to dismiss it as uninteresting. Howev- This ambitious project, carried out jointly at Stanford, SRI and Rockwell Inter- 1 II!

er, it was recognized early on that one would not gain good insight into the national (Nilsson, Rosenschein, Cohen, Moore, Appelt, Buckley, and many oth-
strengths and weaknesses of AOP without writing actual programs. Ir was de- ers) had among its goals the representation of speech acts and connection be-
cided therefore to implement a simple interpreter first, and design more com- tween the intensionallevel and the machine leve!. See discussion of some of the
plex languages and interpreters based on this experience. Ir turned out the de- individual work below.
sign of AGENT-O itself posed some challenges, and we have been surprised by
the diversity of applications that even this simple language admits. Further- Cohen and Levesque's Work on Belief,
more, AGENT-O is designed in a way that suggests obvious extensions; a few Commitment, Intention, and Coordination
are being currently pursued and are descrihed in the final section.
These two researchers (Cohen and Levesque 1997, 1990) have also investigated
The implemented interpreter is documented in Torrance (1991). A second,
more complex interpreter was designed and implemented in collaboration with the logical relationships between several modalities such as belief and choice.
the Hewlett Packard Corporation. Although they have not approached the topic from a programming-language

284 SHOHAM , AGENT-ORIENTED PROGRAMMING 285

formation into action commands. Genesereth uses the term "agents" so as to in-
perspective as I ha ve, they too have been interested in speech acts and mental .~~ 1
elude also low-level finite-automaton-like constructs. AOP's structure of mental
state as building blocks for coordination and analysis of behavior. Their work
has its roots in earlier work in natural language understanding by Allen, state is consistent with Genesereth's deelarative regime but is not required by it. l' 1

Cohen, and Perrault (Allen 1983; Cohen and Perrault 1979). Despite some simi- . I

larities, crucial differences exist between the mental categories employed by Plan Representation and Recognition , I

Cohen and Levesque and ours. ~ I
Work on plan representation and recognition by Kautz, Pollack, Konolige, Lit-
~ I
man, Allen, and others (e.g., Kautz 1990, Litman and Allen 1990, Pollack 1990, 11'
Contract Nets and Bratman 1987) also addresses the interaction between mental state and ac- I

AOP shares with early work on contract nets (Smith 1980) the computational tion, but it is usually concerned with finer-grained analyses, involving the actual b
rol e of contracts among agents. However, the similarity ends there. Contract representation of plans, reasoning limitations, and more complex mental no- iu
,
nets are based on broadcasting contracts and soliciting bids, as opposed to the tions such as goals, desires, and intentions. ~
intimate communication in AOP. Contract nets had no other notion of mental ~
sta te, no range of communicative speech acts, nor any aspect of the asyn- Nilsson's Aetion Nets
~
chronous, real-time design inherent in AOP. ACTNET is a language for eomputing goal-achieving aetions that depends dy-
~
namieally on sensory and stored data. The ACTNET language is based on the ~~
Situated Automata
coneept of action networks. An aetion network is a forest of logical gates that
Rosenschein and Kaelbling's situated automata (Kaelbling 1988; Rosenschein seleet aetions in response to sensory and stored data. The conneetion to AOP, al-
and Kaelbling 1986; Rosenschein 1985) is relevant in connection with the pro- beit a weak one, is that some of the wires in the network originate from
cess of agentification. We adopt their idea of decoupling the machine language database items marked as "beliefs" and "goals." The maintenanee of these
from the programmer's intensional conceptualization of the machine, but differ databases is not the job of the aetion neto
on the specific details.

Coordination Summary of Results and Ongoing Research
Several researchers have been concerned with the process of coordination in
Work on mental state is proeeeding on different fronts. Here are some pointers
modern environments. For example, as a part of their more global project,
Winograd and Flores have developed a model of communication in a work en-
vironment (Winograd and Flores 1986). They point to the fact that every con- .
to ongoing researeh:
In Moses and Shoham (1993) we provide some results on the eonneetion
versation is governed by some rules, which constrain the actions of the partici-
pants: a request must be followed by an accept or a decline, a question by an
answer, and so on. Their model of communication is that of a finite automaton,
.
.
between knowledge and (one kind of) belief.
Thomas (1993) tackles the notions of eapability, plan, and intentions.
In Lamarre and Shoham (1994) we argue for the three-way distinction be-
1
,1

with the automaton states corresponding to different states of the conversation. tween knowledge, certainty, and belief. 1
This is a macro theory, a theory of societies of agents, in contrast to the micra
theory of AOP. In related work, Malone and his associates are aiming towards a
. Brafman and Tennenholtz (1992) layout a framework in whieh beliefs,

general theory of coordination, drawing on diverse fields such as computer sci-
ence and economics (Malone 1991).
. preferenees, and strategy are maintained in a form of "rational balance."
Del Val and Shoham (1994) argue that the properties of belief update
should be derived methodically from a theory of aetion and that doing so
reveals some limitations of the KM postulates.
Informable Agents
Genesereth's work on informable agents (see Genesereth chapter, also in this vol-
. Del Val and Shoham (1994) propose to reduce the notion of belief revision
to that ofbelief update, and thus also to theories of aetion.
ume). Genesereth's interest lies primarily in agents containing deelarative knowl-
edge that can be informed of new facts and that can act on partial plans. In this
. In Shoham and Cousins (1994) we provide an initial survey of logies of
mental state in AL
connection, he has investigated also the compilation of declarative plans and in-
I
!

286 SHOHAM 111

II~I
In parallel with the logical aspects of action and mental sta te, we ha ve investi- that are made up of a group of simpler agents. If we define the beliefs of this composite
~
gated algorithmic questions: agent as the "common beliefs" of the individual agents and the commitments of the com-

· We have proposed a specific mechanism for tracking how beliefs change
posite agents as their "common commitments," then the mental attitUdes of the group
are indeed inherited by the individua!.

· over time, called temporal beliefmaps (Isozaki and Shoham 1992).
In Brafman, Latombe, and Shoham (1993) and Brafman et al. (1993) we
show that, similar to distributed systems, the formal notion of knowledge
3. In this respect our motivation here deviates from that of philosophers. However, 1 be-
lieve there exist sufficient similarities to make the connection between Al and philosophy
mutually beneficia!.
11
4. Cohen and Levesque (1990), for example, propose to reduce the notion of intention to 11
can be applied to algorithmic robot motion planning. Recently, we proposed
those of goal and persistence. Their pioneering work introduces mental categories that
knowledge complexity as a useful general complexity measure in robotics, are different from ours. The two frameworks share the essential view of belief and time.
with an application to automating the distribution of robotic algorithms. They each introduce modalities absent from the other: obligation and capability in our
We have recently begun contemplating the role of agents in the context of framework, goals and intentions in theirs. However, even two notions that at first appear
digital libraries, whether or not they are of the AOP variety. We ha ve so far tObe similar-such as our "decision" and their "choice"-turn out tObe quite different.
conducted one experiment: 5. The term choice is somewhat ambiguous; 1 discuss various senses of choice later.
· There is an experiment to deploy adaptive agents that perform automated
browsing of the W orld Wide Web on behalf of the user.
6. However, an early design of agent programs by Akahani was entirely in the style of
logic programming; in that framework program statements themselves were indeed log-
ical sentences.
Finally, we are .interested in how multiple agents can function usefully in the
presence of other agents. In particular, we are interested in mechanisms that
References
minimize conflicts among agents and ha ve been investigating the utility of so-

·
ciallaws in computational settings:
In Shoham and Tennenholtz (1992) we propose a general framework for
representing sociallaws within a theory of action and investigate the com-
Agre, P., and Chapman, D. 1987. PENGI:An Implementation of a Theory of Activity. In
Proceedings of the Sixth National Conference on Artificial Intelligence, 268-272. Menlo
Park, Calif.: American Association for Artificial Intelligence.
putational complexity of automatically synthesizing useful sociallaws. We Allen, J. F. 1983. Recognizing Intentions from NatUral Language Utterances. In Compu-
also study a special case of traffic laws in a restricted robot environment. tational Models of Discourse, eds. M. Brady and R. e. Berwick, 107-166. Cambridge,
· In Shoham and Tennenholtz (1995) we study ways in which such conven-
tions emerge automatically in a dynamic environment. Early results were
Mass.: MIT Press.
Austin, J. L. 1962. How to Do Things with Words. Cambridge, Mass.: Harvard University
Press.
reponed on in Shoham and Tennenholtz (1992).
· In Kittock (1994), he refines these results to take into account the topology
of the agent network and the existence of asymmetric interactions among
Brafman, R., and Tennenholtz, M. 1994. Belief Ascription and Mental-Level Modeling.
In Proceedings of the Fourth Intemational Conference on PrincipIes of Knowledge Represen-
tation and Reasoning (KR'94). San Francisco, Calif.: Morgan Kaufmann.
agents. Brafman, R. 1.; LatOmbe, J.-C.; and Shoham, Y. 1993. Toward Knowledge-Level Analy-
sis of Motion Planning. In Proceedings of the Eleventh National Conference on
Notes Artificial Intelligence, 670-675. Menlo Park, Calif.: American Association for Artificial
Intelligence.
1. Elsewhere, I discuss how the gradual elimination of animistic explanations with the
increase in knowledge is correlated very nicely with both developmental and evolution- Brafman, R. 1.; Latombe, J.-e.; Moses, Y.; and Shoham, Y. 1993. Knowledge as a Tool in
ary phenomena. In the evolution of science, theological notions were replaced over the Motion Planning under Uncertainty. In Theoretical Aspects of Reasoning about Knowl-
centuries with mathematical ones. Similarly, in Piaget's stages of child development, edge: Proceedings of the Fifth Conference (TARK 1994). San Francisco, Calif.: Morgan
there is a clear transition from animistic stages around the ages of 4-6 (when, for exam- Kaufmann.
pie, children claim that clouds move beca use they follow us around) to the more mature Bratman, M. E. 1987. Intention, Plans, and Practical Reason. Cambridge, Mass.: Harvard
later stages. University Press.
2. There is one more dimension to the comparison, which 1 omitted from the table, and Brooks, R. A. 1986. A Robust Layered Control System for a Mobile Robot. IEEE ¡oumal
it regards inheritance. Inheritance among objects is toda y one of the main features of of Robot Automation 2(1): 14-23.
OOP, constituting an attractive abstraction mechanism. 1 have not discussed it since it is
Cohen, P. R., and Levesque, H. J. 1990. Intention Is Choice with Commitment. Artificial
not essential to the idea of OOP, and even less so to the idea of AOP. Nevertheless a par-
Intelligence 42(3): 213-261.
allel can be drawn here, too. In OOP, specialized objects inherit the methods of more
general ones. One analogous construct in AOP would be group agents, that is, agents Cohen, P. R., and Levesque, H. J. 1997. Rational Interactions as the Basis for Communi-

288 SHOHAM
A(;ENT-ÜRIENTED PRO<;RAMMING 289

cation. In Intentions in Communication, eds. P. R. Cohen, l. Morgan, and M. E. Pollack. so, 415-424. San Francisco, Calif.: Morgan Kaufmann.
Cambridge, Mass.: MIT Press. Forthcoming. Litman, D. l., and Allen, l. F. 1990. Discourse Processing and Commonsense Plans. In
Cohen, P. R., and Perrault, C. R. 1979. Elements of a Plan-Based Theory of Speech Acts. Intentions in Communication, eds. P. R. Cohen, l. Morgan, and M. E. Pollack. Cambridge,
Cognitive Science 3:177-212. Mass.: MIT Press.
Del Val, A., and Shoham, Y. 1994. Deriving Properties of Belief Update from Theories McCarthy, l. 1979. Ascribing Mental Qualities to Machines, Technical Report Memo,
of Action. Joumal 01Logic, Language, and Inlormation 3(2): 81-119. 326, Al Lab, Stanford University.
Del Val, A., and Shoham, Y. 1996. A Unified View of Belief Revision and Update. Jour- Malone, T. W. 1991. Toward an Interdisciplinary Theory ofCoordination, Technical Re-
nal 01Logic and Computation 4(5): 797-810. port, CCS 120, Sloan School of Management, Massachusetts Institute ofTechnology.
Dennett, D. C. 1987. The Intentional Stance. Cambridge, Mass.: MIT Press. MCC. 1990. Proceedings of the Tenth International Workshop on Distributed Artificial
Drummond, M. 1989. Situated Control Rules. In Proceedings olthe First Intemational Intelligence, Technical Report ACT-AI-355-90, MCC, Austin, Texas.
Conlerence on Knowledge Representation and Reasoning, eds. R. l. Brachman and H. l. Mitchell, T. M. 1990. Becoming Increasingly Reactive. In Proceedings of the Eighth Na-
Levesque, 103-113. San Francisco, Calif.: Morgan Kaufmann. tional Conference on Artificial Intelligence, 1050-1058. Menlo Park, Calif.: American
Association for Artificiallntelligence.
Ferber, l., and Carie, P. 1990. Actors and Agents as ReRective Concurrent Objects: A
Mering IV Perspective. In Proceedings of the Tenth International Workshop on Dis- Moses, Y., and Shoham, Y. 1993. Belief as Defeasible Knowledge. Joumal 01Artificial In-
tributed Artificial Intelligence, Technical Report ACT-AI-355-90, MCC, Austin, Texas. telligence 64(2): 299-321.
Genesereth, M. R. 1989. A Proposal for Research on Informable Agents, Technical Re- Pollack, M. E. 1997. Plan as Complex Mental Attitudes. In Intentions in Communica-
port, Logic-89-4, Computer Science Department, Stanford University. tion, eds. P. R. Cohen, l. Morgan, and M. E. Pollack. Cambridge, Mass.: MIT Press.
Forthcoming.
Grice, P. 1989.Studies in the Ways olWords. Cambridge, Mass.: Harvard University Press.
Rosenschein, S. l. 1985. Formal Theories of Knowledge in Al and Robotics, Technical
Halpern, l. Y., and Zuck, L. D. 1987. A Little Knowledge Goes a Long Way: Simple
Report 362, SRI International, Menlo Park, California.
Knowledge-Based Derivations and Correctness Proof.~for a Family of Protocols. In Pro- 1"1
ceedings of the Sixth ACM Symposium on Principies of Distributed Computing, Rosenschein, S. l., and Kaelbling, L. P. 1986. The Synthesis of Digital Machines with
269-280. New York: Association ofComputing Machinery. Provable Epistemic Properties. Paper presented at Conference on Theoretical Aspects of
Reasoning about Knowledge, 19-22 March, Monterey, California.
Hayes-Roth, B.; Washington, R.; Hewett, R.; Hewett, M.; and Seiver, A. 1989. Intelligent
Monitoring and Control. In Proceedings of the Eleventh International loint Conference Schneider, F. 1987. Understanding Protocols for Byzantine Clock Synchronization,
on Artificial Intelligence, 243-249. Menlo Park, Calif.: International loint Conferences Technical Report, Computer Science Department, Cornell University.
on Artificial Intelligence.
Searle, l. R. 1969. Speech Acts: An t.ssay in the Philosophy 01 Language. Cambridge, U.K.:
Hewitt, C. 1990. Toward Open Information Systems Semantics. In Proceedings of the Cambridge University Press.
Tenth International Workshop on Distributed Artificial Intelligence, Technical Report Shoham, Y. 1993. Agent-Oriented Programming. Joumal 01 Artificiallntelligence 60(1):
ACT-AI-355-90, MCC, Austin, Texas. 51-92.
Hewitt, C. 1977. Viewing Control Structures as Patterns of Passing Messages. Artificial Shoham, Y. 1991. Varieties of Contexto In Artificiallntelligence and Mathematical TheOlY
Intelligence 8:323-364.
olComputation, 393-408. San Diego, Calif.: Academic. I
1",
Isozaki, H., amI Shoham, y 1992. A Mechanism for Reasoning about Time and Belief. Shoham, Y. 1989. Time for Action. In Proceedings of the Eleventh International loint
In Proceedings of the Conference on Fifth-Generation Computer Systems, 694-701. Conference on Artificial Intelligence, 954-959. Menlo Park, Calif.: International loint
Amsterdam: lOS Press.
Conferences on Artificial Intelligence.
Kaelbling, L. P. 1988. Goals as Parallel Program Specifications. In Proceedings of the Shoham, Y., and Cousins, S. 1994. Logics of Mental Attitudes in Al: A Very Preliminary
Seventh National Conference on Artificial Intelligence, 60-65. Menlo Park, Calif.: Survey. In Advances in Knowledge Representation and Reasoning, eds. G. Lakemeyer and
American Association for Artificial Intelligence. B. Mebel, 296-309. New York: Springer Verlag.
Kautz, H. A. 1990. A Circumscriptive Theory ofPlan Recognition. In Intentions in Com- Shoham, Y, and Tennenholtz, M. 1992. Emergent Conventions in Multi-Agent Systems:
munication, eds. P. R. Cohen, l. Morgan, and M. E. Pollack, 60-65. Cambridge, Mass.:
MIT Press. Initial Experimental Results and Observations. In Proceedings olthe Conlerence on Prin-
CIpies01Knowledge Representation and Reasoning, 225-23 I. San Francisco, Calif.: Morgan
Kaufmann.
Kittock, l. E. 1994. The Impact of Locality and Authority on the Emergence ofConven-
tions. In Proceedings of the Twelfth National Conference on Artificial Intelligence, Shoham, Y., and Tennenholtz, M. 1995. Computational Social Systems: OfRine Design.
420-425. Menlo Park, Calif.: American Association for Artificial Intelligence. Joumal 01Artificiallntelligence 73(1-2): 231-252.
Lamarre, P., and Shoham, Y. 1994. Knowledge, Certainty, Belief, and Conditional- Smith, R. G. 1980. The Contract Net Protocol: High-Level Communication and Control
ization. In Proceedings 01 the Fourth Intemational Conlerence on Principies 01 Knowl- in a Distributed Problem Sol ver. lEEr: Transactions 01 Compute/' Science 29(12):
edge Representation and Reasoning (KR'94J, eds. l. Doyle, E. Sandewall, and P. Toras- 1104-1113.

-

290 SHOHAM

I

Thomas, B. 1993. PLACA,An Agent-Oriented Language. Ph.D. diss., Depanment of
Computer Science, Stanford University.

Torrance, M. 1991. The AGENTO Programming Manual, Technical Repon, Depanment CHAPTER 14
of Computer Science, Stanford U niversity.

N.J.: Ablex. T., and Flores, F. 1986. Understanding Computers and Cognition. Norwood,
Winograd,
.n
KQML as an Agent Ir
11
Cornrnunication Language, "
,Ii

1,
Tim Finin, Yannis Labrou, &¡ames Mayfield

111

t is doubtful that any conversation about agents will result in a consensus on
I~
1 the definition of an agent or of agency. From personal assistants and "smart"
interfaces to powerful applications, and from autonomous, intelligent enti-
Ir

ties to information retrieval systems, anything might qualify as an agent these
days. But, despite these different viewpoints, most would agree that the capacity
,l~
for interaction and interoperation is desirable. The building block for intelligent
IIJ
interaction is knowledge sharing that includes both mutual understanding of
knowledge and the communication of that knowledge. The importance of such
communication is emphasized by Genesereth, who goes so far as to suggest that
an entity is a software agent if and only if it communicates correctly in an agent
communication language (Genesereth and Ketchpel 1994). After all, it is hard to
picture cyberspace with entities that exist only in isolation; it would go against
our perception of a decentralized, interconnected electronic universe.
How might meaningful, constructive, and intelligent interaction among soft-
ware agents be provided? The same problem for humans requires more than the
knowledge of a common language such as English; it also requires a common
understanding of the terms used in a given contexto A physicist's understanding
of velocity is not the same as that of a car enthusiast's,2 and if the two want to
converse about "fast" cars, they have to speak a "common language." Also, hu-
mans must resort to a shared etiquette of communication that is a result of soci-
etal development and that is partially encoded in the language. Although we are
not always conscious of doing so, we follow certain patterns when we ask ques-
tions or make requests. Such patterns have common elements across human lan-
guages. Likewise, for software agents to interact and interoperate effectively re-
quires three fundamental and distinct components: (i) a common language, (ii) a
common understanding of the knowledge exchanged, and (iii) the ability to ex-

Software Agent Oriented Programming

More Related Content

Similar to Software Agent Oriented Programming (20)

Recently uploaded (20)

Software Agent Oriented Programming