• nothing in that title would look out of place in the programme of any
  agentic-AI venue of today
  • autonomy, agents, multiplicity, a world to act in
• so, when was it written?

• the first European workshop of the field
• the name of the field, at the time, was distributed artificial
  intelligence (DAI)
• thirty-seven years ago, the research question was already written in
  full

• it is not a survey
  • no chronological march: chronology is the substrate; storylines are
    the figure
• it is a claim with evidence
  • every historical beat enters not because it happened, but because
    apparently it is needed again now

• the decoupling is not this speaker’s invention
  • which is precisely what makes it safe to build on
• the mainstream has just discovered that the two axes come apart
  • something the agent community has known since it first had to
    define “agent” at all

• agency is the plane, with two independent axes
  • autonomy
    • who is in control, and of what
  • intelligence
    • what competence is brought to bear
• encapsulation is not a third axis
  • it is the engineering discipline that makes agency constructible
  • it is the reason the abstraction is called an “agent”, and not a
    “process”, an “object”, or a “model”

• this tutorial does not open with a block of definitions
  • unlike many others I have given over the decades
• each notion is introduced when the history forces it, then re-used
  • encapsulation—Act I, and again in Act III
  • autonomy—Act II, and again in Act IV
  • intelligence—Act IV, and again in Act V
• the accumulation is the structure of the argument

• before the name
  • DAI, 1980s–1996
• the theory years
  • 1994–2001
• the field
  • 2002–2010
• the quiet decade
  • 2010–2022
• agentic AI
  • 2022–
• the next decade

• everything the present moment calls new was in the problem
  statement of distributed artificial intelligence (DAI)
  • that intelligence is distributed
  • that it is social
  • that it is situated
  • that a single reasoner is the wrong unit of account
• the field did not lack the idea
  • it lacked the machine

• Modelling Autonomous Agents in a Multi-Agent World (MAAMAW),
  ten editions
  • 1989 Cambridge, 1990, 1991—published by Elsevier / North-Holland
    as Decentralized A.I. 1, 2, 3

    [Demazeau and Müller, 1990] [Demazeau and Müller, 1991] [Werner and Demazeau, 1992]

  • from 1992 the Springer LNAI run: 830, 957, 1069, 1038, 1237, 1647

    [Castelfranchi and Werner, 1994] [Castelfranchi and Müller, 1995] [Perram and Müller, 1996] [Van de Velde and Perram, 1996] [Boman and Van de Velde, 1997] [Garijo and Boman, 1999]

  • 1992 is Artificial Social Systems, edited by Castelfranchi and Werner
• the European half of a field that also had North American DAI
  workshops and the Asian MAAC workshops

• 1995—International Conference on Multi-Agent Systems (ICMAS),
  San Francisco, over 500 participants
  • itself the merger of the European MAAMAW workshops, the Asian
    MAAC workshops, and the North American distributed artificial
    intelligence (DAI) workshops
• four editions—San Francisco 1995, Kyoto 1996, Paris 1998, Boston
  2000
• the field was already organising itself out of three continents, before
  it had a conference
• and it would merge again in 2002
  • see Act III

• Bond & Gasser (eds.), Readings in distributed artificial intelligence,
  Morgan Kaufmann, 1988[Bond and Gasser, 1988]
• O’Hare & Jennings (eds.), Foundations of distributed artificial
  intelligence, Wiley, April 1996[O’Hare and Jennings, 1996]
  • in the “Sixth-Generation Computer Technology” series
    • which is itself a period detail worth noticing
• Weiss (ed.), Multiagent Systems: A Modern Approach to distributed
  artificial intelligence, MIT Press, 1999[Weiss, 1999]
• a field with a canon is a field that expects to last

• read that list again, and say which decade it comes from

• item for item, a 1996 research programme and a 2026 product
  datasheet

• DAI posed an AI question
• but the answer it converged on – the agent – is first and foremost a
  software-engineering answer
• that observation is what this whole tutorial rests on

• this is Parnas’s question, asked one level up
• and it is still open in 2026
  • which is the whole point

• agents work as the basic bricks for the engineering of intelligent
  systems

• we will come back to encapsulation in Act III, with a harder question
  • not what does an agent encapsulate, but what else needs encapsulating

• Agent Theories, Architectures, and Languages (ATAL), 1994–2001
  • eight editions, eight LNAI volumes
    • 890,[Wooldridge and Jennings, 1995b] 1037,[Wooldridge et al., 1996] 1193,[Müller et al., 1997]
      1365,[Singh et al., 1998] 1555,[Müller et al., 1999] 1757,[Jennings and Lespérance, 2000]
      1986,[Castelfranchi and Lespérance, 2001] 2333[Meyer and Tambe, 2002]
  • it started as an ECAI-94 workshop in Amsterdam, 8–9 August 1994,
    edited by Wooldridge and Jennings, 24 papers
• eight book spines in a row: a community building a theory on purpose

• in seven years the field did the conceptual work that the current
  wave is skipping
  • it fully elaborated its conceptual pillars
  • and consequently defined its terms

• every “agentic” definition written since 2023 is a weak-notion
  definition
  • usually with the same four properties, in a different order, with no
    citation

• the Belief, Desire, Intention (BDI) line follows: practical reasoning,
  means-ends reasoning, commitment strategies[Rao and Georgeff, 1995]
• AgentSpeak(L) makes it a programming language[Rao, 1996, Bordini et al., 2006]
• today’s planner / executor / critic loops are strong-notion
  architectures
  • mostly without the hard part: when to (autonomously) reconsider an
    intention

• AI, distributed artificial intelligence (DAI), programming languages,
  parallel & distributed systems, robotics, software engineering
  • six different roads, one destination
• they converge on the notion...
• ...not on the definition
• which is basically the position agentic AI is in today
  • and it is a solvable position, because it was solved once

• not everyone agreed that agency needed representation
  • Brooks
    • the subsumption architecture, intelligence without
      representation[Brooks, 1986, Brooks, 1991]
  • Suchman
    • situated action: plans are resources for action, not programs of
      it[Suchman, 1987]
  • Agre
    • computation and human experience, against the planning
      view[Agre, 1995]
• keep this thread in hand
  • in Act V, “representation is the bottleneck” will land as a return, not as
    a novelty

• monolithic programming
  • nothing is encapsulated
• modular programming
  • unit behaviour is
• object-oriented programming
  • behaviour and state
• agent-oriented programming
  • behaviour, state and invocation

• in object-oriented programming, a message carries data along with
  control
• in agent-oriented programming, a message carries data only
• which is what lets information flow be designed independently of
  control flow

• everything else follows
  • pro-activity
    • “making something happen, rather than waiting for something to
      happen”; autonomous agents are pro-active by definition
  • sociality, situatedness, reactivity, interactivity
• multi-agent systems (MASs) are the only legitimate containers of
  autonomous agents

• etymology
  • Greek autonomia / autonomos, “having its own laws”;
    self-government, self-regulation, self-determination
• philosophy, after Kant
  • “a more demanding notion of autonomy requires not only
    self-determination, but also independence from oneself”
  • the ability to reason on and change one’s own goals, not just one’s
    actions
• the military
  • automatic systems are pre-programmed and “not able to define the
    path according to some given goal”; autonomous ones are
    “self-directed toward a goal... governed by laws and strategies that
    direct their behaviour”
• cars
  • six levels, 0 to 5, “from automatic to autonomous, one step at a
    time”[SAE International, 2021]

• goal-autonomous agent
  • “an agent endowed with its own goals”
• fully socially autonomous agent[Conte and Castelfranchi, 1995]
  • endogenous goals; decides among conflicting goals; adopts outside
    goals only through belief-mediated choice; self-interested
  • “it is impossible to change automatically the beliefs of an agent”
    • which is what protects its cognitive autonomy

• executive autonomy is not enough for real autonomy
• this is the most under-used weapon in the whole story
  • and it is about to become very useful indeed

• an LLM agent reached through a function-call API is invoked
  • its control passes through its own boundary, on every turn
• by the computational definition it is not autonomous
  • in Castelfranchi’s terms, it is executively autonomous at best, and
    motivationally null
• that is a precise critique, not a complaint
  • and it is available only because the definitions were done thirty years
    ago

• Autonomous Agents (ACM), 1997–2001
  • Marina del Rey and after
• PRIMA, from 1998
  • the Pacific Rim
• AgentLink, from November 1998
  • “The European Network of Excellence for Agent-Based Computing”
• European Agent Systems Summer School (EASSS), from 1999
  • still running
• COORDINATION, from 1996
  • models and languages for (agent) interaction[Ciancarini et al., 1999]
• WOA, from 2000[Mascardi and Omicini, 2026]
  • Dagli Oggetti agli Agenti, and it is still running too

• this is the conceptual centre of gravity of this tutorial
  • everything in Act V is measured against it

• three series end, one begins
  • the 6th Autonomous Agents
  • the 5th ICMAS
  • the 9th ATAL
• they become the 1st Autonomous Agents and Multi-Agent Systems
  (AAMAS)
  • general chairs: Gini and Ishida
  • programme chairs: Castelfranchi and Johnson
  • organisation chairs and treasurers: Omicini and Zambonelli
• we can adopt this event as the conventional date when MASs stopped
  being a topic and became a discipline

• three foundational lines stop at 2001, one starts in 2002
• around it, an entire workshop ecology appears and persists
  • AOSE 2000–2012, ESAW 2000–2009, PROMAS and DALT
    2003–2012
  • E4MAS 2004–2014, COIN 2006–2017, ESOA 2003–2006, SASO
    2007–2019
  • EUMAS from 2003, WOA from 2000, EMAS from 2013, EXTRAAMAS
    from 2019
• and AIxIA runs across the whole picture, from 1988
  • the venue we are speaking at is as old as the story

• an FP6 Coordination Action, 2004–2005, coordinated by Peter
  McBurney, Liverpool
• Work Package 5, the Technical Forum, was mine
• three forums
  • TF1—Rome, 30 June – 2 July 2004, over 100 researchers
  • TF2—Ljubljana, April 2005
  • TF3—Budapest, September 2005
• outputs as special issues in KER 20,[Omicini and McBurney, 2005] Informatica
  29[Omicini et al., 2005] and 30,[Omicini et al., 2006a] TAAS[Petta et al., 2007] and
  JAAMAS[Omicini and McBurney, 2008]

• the nine thematic groups of Technical Forum I
  • AOSE,[Bernon et al., 2005a] PROMAS,[Dastani and Gómez-Sanz, 2005]
    SELFORG,[Di Marzo Serugendo et al., 2005] TRUST, LEA, NETAGENTS,
    HEALTHCARE, IIA4WE, BIOAGENTS[Luck and Merelli, 2005]

• methodology, programming, self-organisation, trust, learning,
  agents on the network
• a 2004 snapshot that reads as a perfectly-believable 2026
  workshop call

• Weiss (ed.), MIT Press, 1999[Weiss, 1999]
• Wooldridge, Wiley, 2002[Wooldridge, 2002]
  • 2nd edition 2009[Wooldridge, 2009]
• Vlassis, Morgan & Claypool, 2007[Vlassis, 2007]
• Weyns, Springer, 2010[Weyns, 2010]
• Dastani, Hindriks & Meyer (eds.), Springer, 2010[Dastani et al., 2010]
• Ossowski (ed.), Agreement Technologies, Springer, 2013[Ossowski, 2013]
• textbooks, a summer school, a conference, a journal, a European
  network
  • that is what a discipline looks like

• the field then solved a problem that agentic AI has not yet noticed it
  has

• the environment as a first-class abstraction in MASs[Weyns et al., 2007]
  • with its own abstractions, its own lifecycle, its own
    engineering[Weyns and Holvoet, 2005]
• E4MAS, 2004–2014, is a workshop series about exactly this

• the lineage is explicit and borrowed on purpose
  • Activity Theory
    • every human activity is mediated by mediating artefacts[Vygotskiĭ, 1978]
  • Distributed Cognition
    • cognition is distributed across individuals, tools and artefacts[Kirsh, 1999]
  • cognitive artefacts[Norman, 1992]
    • “those artificial devices that maintain, display, or operate upon
      information in order to serve a representational function and that
      affect human cognitive performance”

• artefacts become a way of programming a MAS,[Ricci et al., 2006] and a way
  of engineering its environment[Viroli et al., 2005, Ricci and Viroli, 2005]
• CArtAgO follows as the infrastructure[Ricci et al., 2009]
• and the meta-model is consolidated in JAAMAS in 2008[Omicini et al., 2008]

• artefacts are transparent, predictable, reactive, situated,
  non-autonomous
• artefacts have no goals
  • they have a function, and, when used, a destination
  • a destination is “an external goal attached to an artefact by an agent,
    in the act of using it”[Conte and Castelfranchi, 1995]
• artefacts do have an interface
  • which is precisely what agents do not have

• AOSE produced methodologies[Bernon et al., 2005b]
  • with phases, abstractions, traceability
  • Gaia,[Zambonelli et al., 2003] SODA,[Omicini, 2001] Tropos[Bresciani et al., 2004]
• SODA extended the methodology to the environment[Molesini et al., 2006]
  • a roadmap to artefacts: design the workspace, not only the society
• organisations get their own abstractions too,[Hübner et al., 2007] and a
  platform that puts the three together[Boissier et al., 2013]

• remember this slide
  • the absence of its 2026 counterpart is the largest open gap in the coda

• coordination
  • a medium that can be programmed[Omicini and Zambonelli, 1999]
• organisation
  • roles, groups, responsibilities, power
• norms and institutions
  • what may be done, and by whom
• the environment itself
  • workspaces, artefacts, tools
• the agent abstraction alone was never going to be enough
  • and the field knew that, and said so, twenty years ago

• while mainstream AI went all-in on learning,[LeCun et al., 2015, Vaswani et al., 2017]
  the agent community kept the abstractions alive and kept
  engineering them
• this act is short, and it is honest about the asymmetry of attention
• the abstraction that the deep-learning decade offered for an
  intelligent system was “the model”
  • no autonomy, no environment, no society, no encapsulation
  • one function, end to end

• PROMAS + AOSE + DALT converge into Engineering Multi-Agent
  Systems (EMAS), from 2013
  • three workshops become one, exactly as three conferences had in
    2002
• AGERE!, 2011–2021, keeps the actors-and-agents conversation
  going with the programming-languages community
• a field that consolidates rather than expands is a field under less
  attention
  • not a field with less to say

• self-organisation and pervasive systems[Di Marzo Serugendo et al., 2006]
  • SAPERE, and a decade of self-organising
    coordination[Zambonelli et al., 2015, Omicini and Viroli, 2011]
  • stigmergy
    • coordination through the medium, named in 1959 for
      termites[Grassé, 1959, Parunak, 2006]
• MAS-based simulation
  • as a method, with validation obligations[Herrler and Klügl, 2006]
• norms, institutions, socio-technical systems—COIN, 2006–2017
• emergence is engineered through the environment, not hoped for
  from population size

• explainable artificial intelligence
  (XAI)[Rosenfeld and Richardson, 2019, Barredo Arrieta et al., 2020]
• EXTRAAMAS from 2019, xAI 2023–2024, and the EXPECTATION
  line
• symbolic knowledge extraction and injection[Ciatto et al., 2024]
  • a principled alternative to reading a trace and calling it an explanation
• this is where intelligence re-enters as a design question rather than a
  performance metric
  • not “how well does it score” but “what about it can be inspected”

• a Turing machine is a heteronomous system; an agent is not
  • which is the deepest available statement of why “agent” is not a
    synonym for “program”

• increasing autonomy is “an evolutionary shift in the
  system–environment relationship”
  • interactive autonomy
    • direct environmental influences gradually reduced
  • constitutive autonomy
    • self-referential intrinsic functions generated and enhanced
• autonomy is not autarchy
  • it is graded, and it is relative

• intelligence is a long-term success story
  • roughly four billion years of it
• in the beginning it was not the word
  • prokaryotes with no mind, no reasoning, and no representation
    nonetheless sensed others and their environment
  • “sensing is the most elementary sort of cognition”
• non-explicit intelligence
  • no logic, no reasoning, no representation
  • just an implicit competence curating life by the rules of homeostasis,
    “relentlessly executed autonomously according to no explicit
    directions”

• quorum sensing
  • coordination with neither mind nor representation[Diggle et al., 2007]
  • a biological precedent for stigmergy, found in bacteria
• there is much more intelligence than just rational intelligence

• autonomy and intelligence are not software properties that
  happened to be invented in 1995
• and, how does intelligence connect with autonomy—in general, and
  within agents?

• agentic AI is the field’s vindication and its embarrassment at once
  • vindication, because agents are now the central abstraction of AI,
    exactly as expected
  • embarrassment, because the re-invention is happening without the
    theory
• and part of the reason is a blind spot of our own, which this act will
  concede

• so “agents and action come into play”
• the mainstream arrives at agents by deduction from a limitation
  • which is a lovely historical irony, and worth naming out loud

• a plain LLM is reducible to sequence continuation
  • P(next token|previous tokens)
• an agentic system wraps that model inside a larger, stateful,
  goal-directed, action-taking process
  • ReAct,[Yao et al., 2022] Toolformer,[Schick et al., 2023] AutoGPT,[Richards, 2023]
    planner–executor, memory-augmented[Lewis et al., 2020]
• they all share: iterative control, explicit state, action semantics,
  environmental feedback, adaptation over time
• “from text as output, to text as control signal”

• on one side, the agentic-AI trade
  literature[Gullí, 2025, Ranjan et al., 2025, Biswas and Talukdar, 2025]
  • an AI agent is “a system designed to perceive its environment and
    take actions to achieve a specific goal... an evolution from a standard
    LLM, enhanced with the abilities to plan, use tools, and interact with
    its surroundings”
• on the other, the MAS engineering literature[Collier et al., 2026]
• a different interpretation for the same terms is apparently available,
  and working
• what is the right one? is there a right one?

• resolve it by applying the test, and showing what passes
  • encapsulated control
  • no interface
  • not invocable
• e.g.
  • an agent behind a function-call schema fails all three
  • an agent with its own control loop, its own goals, and no entry point
    passes all three
• this is not gatekeeping
  • it is the difference between a definition and a marketing category
• and we need to understand each and every available and useful
  notion of agent and MAS, not only our own

• a tool needs a function description
  • what to obtain
• distinct from operating instructions
  • how to exploit it over time
• today’s schemas conflate the two

• the A&A account says what an artefact is, what it is for, and how an
  agent relates to it
  • use, selection, construction, manipulation
    • and a destination attached in the act of using
• the schema account says what arguments a function takes
• both are useful: yet, only one supports composition, and provides a
  basis for design

• cognition distributed across artefacts changes the unit of analysis
  • so memory design is a social-system question, not a retrieval
    optimisation

• organisation is an abstraction with its own semantics – roles,
  responsibilities, power – rather than a call graph between prompts
• and regimentation is not regulation: a norm the environment
  enforces is not a norm the agent is asked to respect

• there were methodologies
  • and no equivalent exists for agentic AI
• this is the largest live gap, and it is squarely in AIxIA’s competence

• the A&A line was in part a reaction to that
  • but A&A did not become the standard; 𝒜gent 𝒞ommunication
    ℒanguage (𝒜𝒞ℒ) did
• so when agentic AI needed a theory of tool use, the standard it
  inherited from the agent world had almost nothing to say
  • and tool calling was built from scratch

• planning is not just fluent plan-shaped text
  • PlanBench: by obfuscating the domain vocabulary while preserving
    the logical structure of the plan, performance
    disappears[Valmeekam et al., 2023]
  • ARC-AGI: a stress test for representation formation, not a verdict on
    general intelligence
• the counter-current of Act II returns here, from the other side
  • Brooks said intelligence needs less representation; the benchmarks
    suggest agency needs more explicit representation
• representation is the bottleneck
• do not confuse plan fluency with autonomous planning competence

• multi-agent systems (MASs) are “at the core of the emerging
  Agentic AI field, providing the conceptual and computational
  foundations for building intelligent systems composed of
  interacting autonomous entities”
• two chapters are the forward vectors of this tutorial
  • ch. 12, from symbolic to neurosymbolic (NeSy) systems[Agiollo et al., 2026]
  • ch. 14, AI infrastructure, from gigastructure to edge
    intelligence[Omicini et al., 2026]

• that sentence is the whole diagnosis
• and each of its three clauses is a research agenda

• so, we should look for a research programme, not a victory lap
• my draft here: one agenda per notion, each stated as a problem you
  could work on

• an LLM agent given a goal and left to achieve it is executively
  autonomous and motivationally null
• a levels-of-autonomy scale built on encapsulated control and the
  executive / motivational axis is concrete, publishable, and obviously
  useful
• it is the single most actionable thing this tutorial can hand you

• which representations must be explicit for an agent to be governable?
• and can those representations be induced, rather than
  hand-designed?
• intermediate representations are not implementation details
  • they are what makes bias contestable and agency
    governable[Ciatto et al., 2025]

• which is Act III’s environment turn returning
  • at the scale of the deployment substrate, rather than the workspace
• the environment was a workspace in 2004, and it is a continent-scale
  infrastructure in 2026
• and it is still the same abstraction problem

• the question has not changed since 1989
  • what is the right abstraction to encapsulate autonomy and intelligence in
    a distributed computational system?
• autonomy was defined, and the definition still discriminates
• intelligence turned out to be separable from agency
  • and plural
• encapsulation is the part still being rebuilt
  • quite badly, right now

• these slides
  talks.apice.unibo.it/MasAixia2026

The Agents Journey: Twenty-Five
Years of Multi-agent Systems

• Springer, LNCS State-of-the-Art
  Survey 16395, April
  2026[Mascardi and Omicini, 2026]
• open access

https://link.springer.com/10.1007/978-3-032-22940-3