Agent swarm and agentic AI glossary
By ContextClone, Systown AI LAB · Last updated
A
- Adversarial review
- An orchestration pattern in which one team of agents builds a case and another team tries to break it. Only the claims that survive the challenge are kept in the final answer.See also:Orchestration patternCross-examination
- Agent (AI agent)
- A language model set up to pursue a goal by taking actions. It decides a step, uses a tool such as search or code, looks at the result and decides the next step, until the goal is met or a limit is reached.See also:AgenticTool (tool use)
- Agent fleet
- A team of AI agents managed as a unit, like a fleet of vehicles. The operator chooses which models fill which roles, how many agents run, what each may see and how much the run may spend.See also:Agent swarmHybrid fleet
- Agent swarm
- A group of AI agents that split one job into parts, work on the parts in parallel and combine the results into one answer. It suits wide work made of many independent units.See also:Multi-agent systemAgent fleet
- Agentic
- Describes AI that works toward a goal by choosing its own steps, instead of answering one prompt and stopping. It is a matter of degree: a system is more agentic the more of the decisions it makes itself.See also:Agent (AI agent)Autonomy level
- Assistant (AI assistant)
- An AI system that responds to each request and leaves the next step to the person. It may use tools, but the person drives. Contrast with an agent, which is given a goal and chooses the steps.See also:Agent (AI agent)
- Audit trail
- A complete, ordered record of what a system did: what each agent received, which model it used, what it returned, what each check decided and what it cost. It lets someone else verify a run afterwards.See also:ReceiptReplay
- Autonomy level
- How much an AI system decides without a person. Levels run from suggesting text, through acting on request and acting with approval, to acting alone within fixed limits. The right level depends on what a mistake costs.See also:AgenticHuman in the loop (human gate)
B
- Baseline (single-agent baseline)
- The result of running the same task with one strong agent, under the same budget, tools and source access as a multi-agent system. Without it there is no evidence that the extra agents helped.See also:Equal-budget comparison
- Budget ceiling
- A fixed upper limit on what a run may spend. A ceiling is only real if spending is blocked before it happens. An alert that fires after the money is spent is a report, not a ceiling.See also:Budget reservation
- Budget reservation
- Setting aside the worst-case cost of a model call before making it, against a fixed ceiling. If the reservation does not fit, the call is refused. Afterwards the unused amount is released. This holds even when many agents call at once.See also:Budget ceiling
- BYOK (bring your own key)
- An arrangement in which the customer supplies their own API keys for model providers. The customer pays the provider directly for inference, sees the real cost, and can change provider without changing platform.See also:Vendor lock-inOpenAI-compatible endpoint
C
- Chain
- A fixed sequence of model calls in which the output of each step is the input of the next, such as extract, then summarise. A person defines the steps in advance, and nothing runs in parallel.See also:Workflow
- Claim
- A single checkable statement produced by an agent, such as a date, an amount or a finding, ideally together with the quoted passage that supports it. Claims are the unit that verification accepts or rejects.See also:EvidenceVerification
- Commit gate
- The point at which a worker's proposed result is either accepted and made visible to the rest of the system, or rejected. Nothing passes the gate without meeting the checks defined for it.See also:Verify, then commit
- Conflict (between sources)
- A case where two sources support different answers, such as a contract and its amendment giving different dates. A careful system records the conflict and shows both, instead of silently choosing one.See also:Reconciler
- Context distribution
- Deciding which parts of the available material each agent in a team receives. Poor distribution sends everything to everyone. Good distribution sends each agent what its task needs.See also:Scoped contextContext duplication
- Context duplication
- Sending the same material to many agents, so the same input tokens are paid for once per agent. It is a common reason that multi-agent cost grows with the number of agents.See also:Context distributionToken
- Context window
- The maximum amount of text, measured in tokens, that a model can take into account in one call. It includes instructions, documents, conversation history and the model's own output.See also:Token
- Coordination overhead
- The tokens and time a multi-agent system spends on planning, hand-offs, status messages and merging, as opposed to the work itself. On tasks that do not split cleanly it can exceed the benefit of parallel work.See also:Orchestration
- Cost ledger
- An exact record of every charge in a run: which call, which model, how many tokens and what it cost. It allows spend to be attributed, audited and reconciled with provider invoices.See also:Audit trail
- Cross-examination
- An orchestration pattern in which two different model families answer the same question independently, and a judge from a third settles the disagreements or keeps them open.See also:Orchestration patternModel family
D
- Decision model
- A model built to return a typed decision, such as yes or no with a probability, instead of free text. In agent teams it is used as a judge, because a narrow typed answer is easier to act on and to audit.See also:Judge (verifier)
E
- Equal-budget comparison
- An experiment in which a single agent and a team run the same task under the same total spending ceiling, including planning, retries and verification. It separates the effect of teamwork from the effect of spending more.See also:Baseline (single-agent baseline)
- Error propagation
- What happens when one agent's wrong output is accepted by another agent as a fact and built upon. Each later step makes the mistake look more established and harder to trace.See also:Verification
F
- Fencing token
- A number that increases each time a task is handed to a worker. A result can only be committed with the current number, so a stalled worker that wakes up late cannot overwrite the result of its replacement.See also:Lease
H
- Hallucination
- Output from a language model that is fluent and confident but not supported by its sources or by fact, such as an invented date, quotation or reference. It looks the same as a correct answer, which is why checking is needed.See also:Verification
- Hierarchical decomposition
- An orchestration pattern in which a planner breaks a large goal into a tree of sub-goals. The leaves run in parallel, and each level is verified before its results roll up to the level above.See also:Orchestration patternPlanner
- Human in the loop (human gate)
- A point in an automated process where a person must review or approve before it continues. It belongs where a decision has consequences that the system should not own, such as rejecting a supplier.See also:Autonomy level
- Hybrid fleet
- An agent fleet that mixes models: large and small, open-weight and closed, from different vendors. Typically a few strong models plan and merge, many small ones do the wide work, and judges come from a different family than the workers.See also:Agent fleetModel family
J
- Judge (verifier)
- An agent or model whose only job is to check another agent's result before it is accepted, for example whether a quoted passage supports a claim. A judge works best when it is independent of the worker it checks.See also:Decision modelVerification
L
- Lease
- A time-limited claim on a task. While a worker holds the lease, no other worker may take the task. If the worker stalls, the lease expires and the task can be given to another.See also:Fencing token
M
- Map, verify, reduce
- An orchestration pattern for wide work. Split a large pile into units, give each worker only its unit, check every claim against evidence, and merge the claims that hold into one answer.See also:Orchestration pattern
- Model family
- A line of models from one developer that share architecture and training approach, such as the GPT, Claude, Gemini, Llama or Qwen families. Models in a family tend to share strengths and blind spots.See also:Hybrid fleet
- Multi-agent system
- Software in which several AI agents work on one job, usually with different roles: a planner splits the work, workers do the parts, a verifier checks results and a reconciler merges them.See also:Agent swarm
O
- Observability
- The ability to see what a system did and why from the records it keeps. For agent teams this means knowing what each agent received, produced and cost, not only the final answer.See also:Audit trail
- Open-weight model
- A model whose trained weights are published, so anyone can run it on their own hardware, subject to its licence. It is not the same as open source: training data and code may still be private.See also:Model family
- OpenAI-compatible endpoint
- A model server that accepts requests in the same format as the OpenAI API. Many hosted services and self-hosted servers offer one, which lets software switch models by changing an address and a key.See also:BYOK (bring your own key)Vendor lock-in
- Orchestration
- Coordinating agents at run time: assigning tasks, passing context, handling failures, enforcing budgets and merging results. An orchestrator may follow a fixed workflow or let a planner decide the steps for each job.See also:Orchestration pattern
- Orchestration pattern
- The shape of a collaboration between agents: who works on what, in what order, and who decides. Examples are map-verify-reduce, cross-examination, specialist panel, adversarial review, watchtower and tournament.See also:Orchestration
P
- Planner
- The agent that reads the goal, decides whether and how to split it into tasks, and works out what each worker needs to know. It usually runs on a strong model, because it is called only a few times per job.See also:WorkerReconciler
- Prompt injection
- Text planted in a document, web page or email that a model reads and then obeys as if it were an instruction. In a team the risk grows, because other agents may trust the hijacked agent's output.See also:Scoped context
R
- Receipt
- The record of one committed result: its inputs, model, output, checks and cost. When each receipt includes a hash of the previous one, the chain cannot be altered later without the change being detectable.See also:ReplayAudit trail
- Reconciler
- The agent that merges accepted results into one answer at the end of a run. A good reconciler keeps disagreements between sources visible instead of smoothing them over.See also:PlannerConflict (between sources)
S
- Scoped context
- Giving each agent only the context its task needs and is authorized to see. Rules that apply to every task still travel with it, and what was left out is recorded. It lowers cost and limits exposure.See also:Context distribution
- Specialist panel
- An orchestration pattern in which several reviewers with different briefs, such as legal, financial, technical and safety, read the same scoped material and each report against their own checklist.See also:Orchestration pattern
- Stale context
- Information an agent holds that is no longer current, because a source was replaced or another agent's result was corrected after it was read. Work built on stale context may need to be redone.See also:Watchtower
T
- Token
- The unit in which language models read and write text, roughly a short word or part of a word. Providers charge per token, usually at different rates for input and output.See also:Context window
- Tool (tool use)
- A function a model can call to act outside itself: search, read a file, run code, query a database, send a message. Tools are what turn a model that talks into an agent that does things.See also:Agent (AI agent)
- Tournament
- An orchestration pattern in which many candidate answers are produced cheaply in parallel, then scored and narrowed over several rounds by independent judges until the best remain.See also:Orchestration patternJudge (verifier)
V
- Vendor lock-in
- Dependence on one provider's models, tools or hosting that makes switching costly. It matters in AI because model quality and price change often, and the best choice for a role this year may not be next year.See also:BYOK (bring your own key)OpenAI-compatible endpoint
- Verification
- Checking an agent's result before relying on it. It ranges from mechanical checks, such as whether a quoted passage exists in the source, to a judgement of whether the evidence really supports the claim.See also:Judge (verifier)Verify, then commit
- Verify, then commit
- A rule that a worker's output is only a proposal until it passes verification. Accepted results are committed and become visible to other agents. Failed proposals are retried or refused, never silently kept.See also:Commit gateVerification
W
- Watchtower
- An orchestration pattern in which many sources are re-checked on a schedule. When one changes, only the work that depended on it is redone, instead of the whole job.See also:Orchestration patternStale context
- Worker
- An agent that carries out one task with the context it was given and returns a result with its evidence. Workers are usually many, run in parallel, and use small, fast, inexpensive models.See also:PlannerJudge (verifier)
- Workflow
- A process whose steps a person defined in advance, possibly with branches and loops. A model may carry out each step, but the path is fixed. It is more predictable than an agent and often the better choice when steps are known.See also:ChainAgent (AI agent)
Guides that use these terms
- What is an agent swarm?A definition, how a swarm differs from one model with tools, from a workflow and from a chain, and when it is the right or the wrong shape.
- What is agentic AI?Agentic as a property: goals, tools, memory and levels of autonomy. Agent versus assistant versus workflow, and what the word does not mean.
- What are multi-agent systems and agent fleets?The roles in an agent team, hybrid fleets that mix model families and sizes, single-family versus mixed-family, and seven orchestration patterns.
- What are the challenges of multi-agent AI?Ten real failure modes, each with a plain example: duplicated context, error propagation, coordination overhead, unverifiable output, runaway spend and more.
- How do you make agent swarms reliable?Seven practices anyone can apply, from scoped context to measuring against a single-agent baseline, and how ContextClone implements them.
- Single agent or multi-agent: which should you use?An honest comparison of one agent against a team, what published evidence does and does not show, and a checklist for deciding.
- What does an agent swarm cost, and what is the ROI?Where the money actually goes: people's checking time, not tokens. A three-column worked example, labelled as arithmetic, and a calculator.