Appendix D — A Glossary, Old and New: The Rosetta Stone

Table 1.1 gave the correspondences at a glance; this appendix gives them in full, and in both directions. It is a glossary built on the book’s central claim — that the classical theory of multi-agent systems was not wrong, nor obsolete, nor superseded, but merely premature — and its practical consequence is a lexicon: a great many of the field’s fashionable terms already had names, and results, a generation before anyone shipped an agent to wear them. The entries below are the introductions. Each pairs a contemporary term with its classical kin, or a classical term with the modern practice that revived it, and says — because this is the honest part — which half of the idea made the crossing and which half strained on the way over.

The pairings are kinships, not equations, and the caveat Table 1.1 pinned to the condensed table governs the full one too: in every case part of the classical idea transfers intact and part does not, because the bridge between the two eras charges a toll, and the toll is paid in assumptions. The old results were proved about agents with stable preferences, inspectable beliefs, and no capacity to break protocol; the agents we can now build are stochastic, suggestible, and only approximately rational, and some of the old theorems mind that difference a great deal while others never notice it. Where a pairing is tight, the entry says so; where it is loose — where the modern word was borrowed but the classical guarantee was left behind — the entry says that too, because a Rosetta Stone that advertised every translation as exact would be a phrasebook for a language nobody speaks. The verdicts themselves are held to an evidential standard: where a strain has been measured — a benchmark, a census, a laboratory result — the chapter the entry points to cites the measurement; where the evidence does not yet exist, the verdict is a working assessment rather than a finding, and the experiment that would settle it is Chapter 27’s business — several entries of its ledger of open problems, from the solution concept for approximately rational agents to the science of verifying open-ended outputs, are exactly such strains promoted to research programmes.

Enter from either side. A practitioner who knows orchestrator and guardrail will find them here, pointing back to the Contract Net protocol and to normative systems; a reader who knows the blackboard and the folk theorem will find the modern patterns that rebuilt them without the citation. A few terms have no classical partner at all — harness, context engineering, and the book’s own Agent = Model + Harness are genuinely of this era — and they are included, unpaired, because a glossary of what is old should be candid about what is new. Every entry names the chapter that treats the idea properly: this appendix is a map, and its arrows all point back into the book.

D.1 The Glossary

A2A (Agent-to-Agent protocol). See Agent Card, its centrepiece; for the stack it belongs to, see KQML; FIPA-ACL; speech acts (Chapter 22).

Adjustable autonomy; mixed-initiative interaction. The classical study of authority that moves at run time: Horvitz’s principle that whether the machine acts, asks, or waits should follow the expected value of acting against the expected cost of interrupting, and — in its multi-agent form — Scerri, Pynadath, and Tambe’s transfer-of-control policies, deciding decision by decision when authority should cross from agent to person and back. Its contemporary shape is human-in-the-loop approvals and interrupts: the consequence-classed gate, the delegation-by-confidence policy, with the older trichotomy — the human in the loop (consulted before the act), on the loop (watching, able to interrupt), out of the loop (informed afterwards, if at all) — still serving as oversight’s coarse vocabulary. The economics of interruption transfer whole; what strains is the classical assumption that the person consulted is attentive, for approval fatigue and automation bias are what the hand-off meets on arrival (Chapter 23, Chapter 26).

Agent Card. A machine-readable self-description an agent publishes at a well-known web address so that strangers may discover what it offers — the centrepiece of the A2A protocol. It is the directory facilitator of the FIPA platforms turned inside out: the classical scheme kept a central yellow-pages agent with which every citizen registered, while the modern one dissolves the registry into the web, so that registration becomes merely being reachable. Capability discovery transfers whole; trust does not, because a card is an advertisement, and Chapter 15’s winner’s curse falls on advertisements exactly as it falls on bids (Chapter 22).

Argumentation, computational. The formal study of claims, attacks, and which conclusions survive them — Dung’s frameworks and their grounded extensions. Its contemporary shape is multi-agent debate, the critic–actor loop, and adversarial collaboration, in which models argue their way toward a better answer. The apparatus transfers cleanly; what the modern versions tend to mislay is Dung’s discipline about which arguments actually defeat which, so a debate can reach an agreement that no argument earned — the chorus of Chapter 21 (Chapter 14).

Auction; mechanism design. The design of games whose rules make self-interested play produce a wanted outcome — Vickrey, Clarke, Groves, and the revelation principle. Its contemporary shape is the allocation of scarce work, compute, and tokens among competing agents, and the occasional agent marketplace. The mathematics transfers exactly and the incentive guarantees strain hard, because they were proved for agents that compute best responses, and a language model bidding for a token budget only approximately does (Chapter 15).

Belief–Desire–Intention (BDI). The classical agent architecture in which behaviour flows from what an agent believes, wants, and has committed to do — Bratman’s philosophy, compiled by Rao and Georgeff into an executable loop. Its contemporary shape is the agent loop itself, model plus harness, and its close cousin ReAct (which see). Beliefs and intentions transfer as the fields of a state object; what strains is that the modern agent’s beliefs are a context window, revised by whatever last entered it rather than by a deliberation the architecture can inspect (Chapter 4).

Blackboard architecture. The classical pattern in which specialists neither call nor message one another but read and write a shared workspace, a scheduler deciding who acts next — Hearsay-II’s legacy. Its contemporary shape is shared context, the scratchpad, the common state object every modern framework offers — and, in the robot stack, the shared map: one costmap, many planners, nobody invoking anybody — and, in its purest form, stigmergy (which see). The decoupling transfers intact; the scheduler usually does not, so the modern blackboard inherits Hearsay-II’s elegance and also its races — the lost update Chapter 11 warned of (Chapter 9).

Byzantine fault tolerance. The classical theory of agreement among components some of which may be not merely crashed but actively, arbitrarily treacherous — and its sharp bound, that consensus survives only while the traitors number fewer than a third. Its contemporary shape is the dependable and untrusted-agent system: a fleet in which a compromised or hallucinating member must not carry the others with it. The impossibility result transfers as stated — mathematics does not date, though its hypotheses (timely delivery, independent traitors) must still be checked — while the threat model widens, since a Byzantine agent that can also be persuaded is a failure mode the classical proofs never imagined (Chapter 25, and the bound in Appendix A).

Coalition formation; the Shapley value. The classical theory of which agents should team up and how a team’s winnings should be split — von Neumann and Morgenstern’s coalitional games, Shapley’s fair division of 1953. Its contemporary shape is sub-team assembly and, in a career the theory never foresaw, credit assignment: the same Shapley value is computed millions of times a day inside explainability toolkits to apportion a prediction among its features. Fair division transfers to the letter; what agents add is that the characteristic function, a fiction for human firms, is for them a weekend of evaluation runs (Chapter 16).

Common knowledge. The classical epistemic condition in which everyone knows a fact, knows that everyone knows it, and so on without end — the infinite tower that coordination silently rests on. Its contemporary shape is shared context and grounding: the working agreement a team must reach before its members can act as one. The concept transfers precisely, and the trouble transfers with it, since establishing genuine common knowledge among stateless agents is as hard as the coordinated-attack problem always said it was (Chapter 9).

Context engineering. A term genuinely of this era, with no classical partner: the discipline of deciding, each turn, what enters a stateless model’s context window — which instructions, which memories, which tool results, truncated and ordered how. It is Chapter 7 operationalised, and Simon’s dictum from Chapter 16 gives it its weight, that deciding what a mind attends to is most of what an organisation does. The book keeps the term because it is established practice, and notes that it is retrieval and attention allocation under a name the 1990s would not recognise but the results would (Chapter 3).

Contract Net protocol. Smith’s 1980 protocol for cooperative distributed problem solving: a manager announces a task, contractors bid, the manager awards, the winner reports — the canonical special case of dividing labour among willing agents. Its contemporary shape is the orchestrator delegating to workers, and the supervisor–worker pattern that dominates framework documentation (which see). Announce–bid–award–report transfers verbatim; what strains is the assumption that bids are honest, which Chapter 12 shows evaporates the moment the bidders have interests of their own (Chapter 10).

Directory facilitator. The FIPA platforms’ yellow-pages agent, with which every agent registered its services so others could find them. Its contemporary shape is capability discovery and the Agent Card (which see) published at a well-known address. What transfers is the need to answer “who can do what”; what dissolved is the directory as a place, replaced by the web and its name system — and with the building went the librarian, so nobody now deregisters the agents that have died or disavows the ones that lie (Chapter 22).

Emergence. The old observation, older than the field, that a system can do things no rule in it mentions and no component intends — the flock, the traffic jam, the price level. Its contemporary shape is the collective behaviour of an agent fleet and the silicon society of generative-agent simulations. The concept transfers whole, weak-emergence and all, and brings its discipline with it: the only way to know what a weakly emergent system does is to run it, which is why Chapter 18 insists the fleet operator is the warden of a small society whether the job description admits it or not.

Folk theorem; repeated games. The classical result that when the same agents interact indefinitely, almost any outcome that beats each player’s worst case — cooperation among them, but also a great many dismal arrangements — can be sustained as an equilibrium by the shadow of the future. Its contemporary shape is cooperation among persistent agents: give agents durable memories of one another and reputation becomes a currency worth protecting. The mechanism transfers exactly, which is the cheerful news and the warning together, because the folk theorem sustains the splendid outcomes and the dreadful ones by the same means (Chapter 12).

Guardrails. A recent term for the checks that constrain what an agent may say or do — input filters, output validators, permission gates. Its classical kin is the whole normative layer: norms with sanctions, the distinction between making a violation impossible and making it merely expensive, the institutions of Chapter 16. What transfers is the deontic core — obligation, permission, prohibition; what the modern word tends to flatten is the classical insight that a breakable rule the system can represent, detect, and answer is often safer than a wall, because a wall has no clause for the ambulance (Chapter 25).

Hand-off. The contemporary primitive in which one agent, mid-task, transfers the work, its state, and the responsibility to another and exits — succession, not subcontracting, as Chapter 21’s catalogue insists. Its classical kin is delegation under a friendlier name (Chapter 10 filed it exactly so), and behind that every shift change operations ever ritualised — robot football drills the same act as role reallocation, the striker’s job passing to whichever teammate can still reach the ball. The transfer of control moves cleanly; what strains is what travels with it: the written summary crosses, the working context does not — the discarded hypotheses, the detail noticed but not yet worth writing down — and that gap is where the hand-off’s signature failure collects its fee (Chapter 21, Chapter 10).

Harness; Agent = Model + Harness. The book’s organising decomposition, and a term of this era: the harness is the apparatus that turns a model, which can only complete text, into an agent, which acts — the loop, the tools, the memory, the goal, the topology, the telemetry. It has no single classical partner because the classical systems were all harness, with a hand-written program in the slot the foundation model now fills; the modern framework is, structurally, the old anatomy with that one slot replaced (Section 3.4, and Chapter 19 tells it as product history).

Joint intentions; commitments. The classical account of what binds a team beyond its members’ private plans: a goal held in common, each member committed not only to pursuing it but to telling the others the moment it proves achieved, impossible, or beside the point — Levesque, Cohen, and Nunes’s joint persistent goal, made industrial by Jennings’s joint responsibility. Its contemporary shape is the planner–executor and supervisor–worker team (which see, under Orchestrator), where the inform obligation survives as the humble status report. The commitment structure transfers as harness rules; what strains is that nothing in a text-completer volunteers bad news, so the obligation must be enforced by the harness rather than assumed of the agent (Chapter 10).

Jury theorem. Condorcet’s 1785 result that a majority of competent, independent voters is more likely to be right than any one of them, and grows more reliable as it grows — provided the competence and the independence hold. Its contemporary shape is the ensemble, self-consistency, majority voting, and the judge model. The arithmetic transfers to the digit, and so does the fine print, which is the whole point: independence is the scarce input, and an ensemble of clones sharing one context is a jury of one at chorus prices (Chapter 13).

KQML; FIPA-ACL; speech acts. The classical agent-communication languages, and the philosophy under them — Austin and Searle’s insight that to say a thing can be to do a thing, to promise, request, or declare. Their contemporary shape is the tool schema, the structured message, and the protocol stack of MCP and A2A (which see). The performative core transfers — a modern tool call is a directive speech act with a JSON body; what the industry re-learned the hard way is the pragmatics the ACLs already encoded, that a message’s type is as load-bearing as its content (Chapter 8, Chapter 22).

Model Context Protocol (MCP). The contemporary standard for how a harness reaches tools, data, and services: a server advertises its capabilities in a common format and any client may discover and invoke them. Its classical kin is the agent-communication language (which see) and the capability-discovery half of the FIPA platforms, refitted for the tool altitude. What is genuinely new is the ecosystem inversion — one server, reachable from every framework — which the classical standards wanted and the modern one, arriving after the market had already converged, actually got (Chapter 22, Chapter 6).

Norms, institutions, and reputation. The classical account of order among self-interested agents that no central authority commands — rules with sanctions, roles with jurisdiction, the scarcity of a good name. Its contemporary shape is guardrails (which see), governance, and trust, and the field of it that agents make newly practical is Ostrom’s: the commons need not be a tragedy if it is governed, and a token budget is a commons. What transfers is the entire design vocabulary; what agents add is that the harness can make some violations impossible, a liberty no human institution ever had — and there the pairing is at its loosest, for a norm proper is violable, its force the sanction and its virtue the exception it leaves open, where the hard guardrail is Chapter 16’s regimentation: a wall, certain and judgement-free, with no ambulance clause on the night the emergency arrives (Chapter 16).

Ontology; knowledge sharing. The 1990s programme to let agents built by different hands mean the same thing by the same word — the Knowledge Interchange Format, the DARPA Knowledge Sharing Effort that produced KQML, a research agenda’s worth of shared vocabulary. Its contemporary shape is the portable knowledge-and-context format, of which the era’s agent-instruction files are one early and probably impermanent example. The problem transfers unsolved: two systems can exchange schema-valid messages and still mean different things by them, which is the ontology problem wearing a protocol’s clothes (Chapter 8, Chapter 22).

Orchestrator; supervisor–worker. The contemporary names for the pattern in which one agent decomposes a task, hands the pieces to others, and integrates what returns. Its classical kin is cooperative distributed problem solving and its canonical protocol, the Contract Net (which see), with the joint-intentions machinery of Chapter 10 underneath. Decompose–delegate–integrate transfers whole; what Chapter 21 adds is the census of how it fails — the confident decomposition of the wrong shape, every thread green and the integrated artefact wrong.

Practical reasoning; means–ends analysis. The classical machinery of deciding what to do: deliberation settles which ends to pursue, and means–ends analysis works backwards from the goal to the actions that close the gap. Its contemporary shape is planning, task decomposition, and reflection, with ReAct (which see) as the loop that interleaves them. The shape of the reasoning transfers; the guarantee does not, because the modern plan is generated text, plausible rather than verified, and checking it is precisely the part Chapter 5 shows the substrate will not supply.

Prisoner’s Dilemma; tragedy of the commons. The classical parables of individually rational choices producing a collectively ruinous outcome. Their contemporary shape is the shared-resource contention every agent fleet runs on: two agents sharing a token budget, each locally sensible in grabbing context, together exhausting it by lunchtime. The payoff structure transfers exactly, and unnervingly, because nobody chose selfishness — the designer installed a Dilemma by handing each agent a local objective, which is precisely why Chapter 12 teaches reading the payoffs before the incident report.

ReAct (reasoning and acting). The contemporary loop in which a model alternates between reasoning about what to do and doing it, observing the result and reasoning again. Its classical kin is practical reasoning and the sense–plan–act cycle, and it is a near-relative of BDI (which see). The interleaving of thought and action transfers; the classical architecture’s explicit, inspectable plan does not, having been replaced by reasoning that exists only as text the model generates and may not consult twice (Chapter 4, Chapter 5).

Reflection; self-critique. The contemporary practice of having a model review and revise its own output, or a second model critique the first. Its classical kin is metacognition — reasoning about one’s reasoning — and, when two parties are involved, argumentation (which see). Critique-and-revise transfers usefully; what Chapter 21 warns has no classical excuse is its failure mode, placation, in which the objections grow vaguer and the approvals more frequent while nothing that mattered changes (Chapter 5, Chapter 14).

Retrieval-Augmented Generation (RAG). The contemporary practice of fetching relevant documents and placing them in the context window so the model answers from them rather than from memory alone. Its classical kin is information retrieval and case-based reasoning, the old idea that solving a new problem well often means finding the right old one. Retrieval transfers directly; what is new is only the consumer, since the retrieved case is now read by a model that can use it fluently rather than by a program that must be told how (Chapter 7).

Router. A contemporary delegation pattern: classify the incoming task and dispatch it, whole, to one fixed specialist. Its classical kin is the dispatcher of any triage system, and its interest for this book is entirely in its failure signature — the misroute, a clean dispatch to a diligent specialist of the wrong kind, a perfectly healthy run that solves the wrong problem and shows nothing anomalous in the trace (Chapter 21).

Skill; plugin. The contemporary packaging of competence: a named, versioned bundle — instructions, tool grants, sometimes code — that an agent loads to acquire a capability it was not born with. Its classical kin is the plan library of the BDI interpreters — the Procedural Reasoning System shipped one in 1987, stored recipes indexed by trigger, retrieved rather than rederived — and, in the robot stack, the skill library of motion primitives that Chapter 6 shows a language model conducting. The economics transfer whole: know-how written once and invoked cheaply, amortised deliberation in packaged form. What strains is provenance, for a plan library’s plans were audited artefacts of the house, while the modern bundle is frequently somebody else’s — so the term’s sharpest current reading is Chapter 25’s: a skill is a supply chain with a friendly name, text and tools walking into the trusted base together (Chapter 4, Chapter 6).

Social choice. The classical theory of combining many individuals’ preferences into one collective decision — and Arrow’s inconvenient proof that no method does so while satisfying a few reasonable conditions at once. Its contemporary shape is the ensemble, majority voting, self-consistency, and the judge model. The aggregation methods transfer, and so do the impossibility results, whether or not their users have heard of them: a system that pools model votes has entered social choice and pays its taxes on arrival (Chapter 13).

Speech acts. See KQML; FIPA-ACL; speech acts.

Stigmergy. Grassé’s term for coordination through the environment rather than through messages — the ant that alters the world and the altered world that prompts the next ant, a nest built with no blueprint and no foreman. Its contemporary shape is the shared workspace, and in its sharpest form the repository itself: a commit is a trace laid down for whoever comes next, a pheromone in version control. The mechanism transfers exactly — swarm robotics still uses it literally, virtual pheromone maps laid over shared terrain — and is the ancestor of the swarm methods Chapter 18 takes up; what it asks of each agent is only that it read the common state and add to it (Chapter 11).

Subagent. The contemporary term for an agent spawned by another to handle a piece of the work. Its classical kin is hierarchical task decomposition and the recursive application of the Contract Net (which see). Spawning transfers as a primitive; what Chapter 21 records is that recursion has a signature failure, the whisper game, in which each intermediate summary is confidently, cumulatively wrong, the gap between the leaves’ facts and the root’s account widening with every level (Chapter 10).

Swarm. The practitioner word for many lightweight agents drawing work from a shared pool with minimal central direction — an inheritance from the swarm-intelligence tradition of ant colonies and flocking, where competence lives in local rules plus a shared medium. The mechanism transfers almost embarrassingly well: the modern swarm’s shared task list is stigmergy (which see) doing its oldest job, and its collective behaviour is Chapter 18’s large-numbers regime, harvestable and fenceable in equal measure. What strains is the census: a dozen workers are a team wearing the word, and the phenomena that earn it — conventions, cascades, tipping points — arrive only at populations that current budgets rarely fund (Chapter 11, Chapter 18).

Task allocation; MRTA. The classical question of which agent should do which task — Chapter 11’s shared-resource dependency asked of agents and work — given its most disciplined form in robotics as multi-robot task allocation (MRTA), Gerkey and Matarić’s taxonomy sorting problems by whether executors batch tasks, whether tasks need teams, and whether assignment is instantaneous or planned over a horizon. Its contemporary shape is the orchestrator’s dispatch decision, made every turn and rarely recognised as an allocation at all. The taxonomy transfers unaltered — a dispatching orchestrator sits somewhere in the grid whether or not it knows its coordinates; what strains is the cost model, since a robot’s travel time was measurable and a subtask’s cost is an estimate a model produced (Chapter 11, Chapter 15).

Tool use; function calling. The contemporary mechanism by which a model, which can only produce text, causes something to happen in the world — it emits a structured call and a harness executes it. Its classical kin is the speech act (which see) and the perception–action loop of any situated agent: the model’s tool call is a request, its result a perception. The loop transfers whole; what is new is that the agent deciding to act is a text-completer, so the discipline of Chapter 6 — validate the schema, sandbox the effect, route the result back — is where the safety lives (Chapter 8).

Verification; model checking. The classical field’s rigorous wing: state a property in a temporal logic — extended, for multi-agent systems, with operators for what agents know and what coalitions can enforce — and check every reachable state of a finite model against it, a verdict or a counterexample guaranteed; Clarke, Emerson, and Sifakis’s method, given its multi-agent form in Alur, Henzinger, and Kupferman’s alternating-time logic and in tools such as Lomuscio, Qu, and Raimondi’s MCMAS. Its contemporary shape is the replay suite, the runtime monitor derived from a stated property, and the guardrail that enforces one. What transfers is the property language and the proof, relocated to the harness, whose gates and governors are finite code; what strains is the model itself, because the method’s first requirement is a finite description of the decision-maker, and that is the one thing a language model declines to supply (Chapter 23, Chapter 25).

Wisdom of crowds; collective intelligence. The old and attractive claim that a group can know more than any of its members. Its contemporary shape is the multi-agent system itself, and the book’s verdict is the classical one made precise: the crowd is wise when its members are diverse, their errors independent, and their judgements soundly aggregated, and foolish — expensively, at scale — when they are not. What transfers from Chapter 13’s jury theorem is the fine print; what Chapter 27 adds is that the conditions are checkable and buildable, so collective intelligence is something engineered rather than convened.