All chapters
  1. Start hereOne agent asks. Another answers. The rest is the contract between them
  2. Foundations · 1One model, one tool, two agents: who is who, and who is responsible
  3. Foundations · 2Message passing: sender, recipient, payload, address and correlation
  4. Foundations · 3Topologies: central orchestrator, peer-to-peer and hierarchy
  5. Mechanics · 4Request and reply, or a task: synchronous versus asynchronous exchange
  6. Mechanics · 5Contracts: messages, artifacts, typed schemas, errors and versions
  7. Mechanics · 6The task lifecycle: submitted, working, input-required, completed, failed, canceled
  8. Mechanics · 7Discovery: Agent Cards, skills and how an agent says what it can do
  9. Protocol · 8A2A: the Agent2Agent protocol, part by part
  10. Protocol · 9A2A versus MCP: delegating to an agent versus calling a tool
  11. Protocol · 10Handoff, delegation or shared state: three ways to work together
  12. Production · 11How updates arrive: streaming, push notifications, callbacks, polling and queues
  13. Production · 12Security: identity, delegated scopes, trust across domains, authorization
  14. Production · 13Reliability: retries, idempotency, order, timeouts and compensation
  15. Production · 14Observability: one trace across three agents
  16. Reference · 15Frameworks: A2A SDKs, LangGraph, AutoGen, CrewAI and plain code
  17. Reference · 16Advanced: consensus claims, coordination failure, cost and isolation
  18. ReferenceGlossary
  19. AboutAbout this site and its sources
Reference

Glossary

Every term this site uses, each with what it is, what it is not, and how it shows up in the insurance example.

Every term this site uses, in one place. Each card opens the same drill-down as the underlined phrases in the text: one line, how it shows up in the insurance example, what it is, what it is not, and a concrete case.

Three sentences hold the words apart. A tool is called. An agent is asked. A description is a claim, not a guarantee.

A2A SDKCode you import that builds and parses A2A messages, serves an Agent Card, and exposes or calls the A2A operations over the protocol's bindings.A2AA written specification for how one agent finds another, sends it messages, delegates tasks to it and receives artifacts, so that agents built by different teams can work together.ActorAny participant in the system: an agent, a tool or a person.AddressThe name the transport uses to deliver a message to its recipient.Agent as a toolA pattern in which one tool name, with arguments in and a result out, hides an agent that would otherwise be asked for a task.Agent CardThe document in which an agent says who it is, where it lives, what skills it offers and how to authenticate. A2A's name for it.Agent communicationEverything about the messages agents exchange: their shape, addresses, correlation, task states, discovery, security, delivery and tracing.AgentA program that calls a language model in a loop and runs the tools the model asks for, until the work is done.ArtifactThe result a task produces: a structured, versioned output with its own id, kept after the exchange ends.Assessment agentThe agent that collects the evidence with the three read tools and returns one evidence artifact.AsynchronousAn exchange in which the caller sends the request and goes on; the result arrives later, in a separate message.At-least-once deliveryA delivery promise: every message arrives, and some arrive more than once.Attempt numberA number that says which try of a step this is: 1, then 2 for a retry, inside the same run.AuthenticationChecking that a request really comes from the identity it names.AuthorizationThe decision, per request, whether this identity may take this action on this resource.AutoGenA library from Microsoft Research in which agents exchange typed messages through a runtime, by direct request and reply or by broadcast to topics.Backward-compatible changeA change after which a consumer written for the old version can still read what the producer sends.BlackboardA shared store that several agents watch, each adding what it can when the board shows something it can act on.Blocking callA call that holds the caller's thread, connection or loop until the answer arrives or the wait is abandoned.CallbackA way for the callee to reach the caller later, given by the caller with the request.CancellationThe caller's request to stop a task before it finishes.CapabilityA protocol feature an agent declares in its Agent Card, such as streaming updates or push notifications.Cascade failureOne failure that causes others: slowness leads to timeouts, timeouts to retries, retries to more load and more slowness.Causal timelineThe trace drawn as bars against time, children under parents, so cause and duration show in one view.Central orchestrationA topology with one agent or program in the middle that sends every request and receives every reply.ChoreographyA way to coordinate where each participant reacts to messages or events from the others, and no one directs the flow.CompensationWhen part of the work cannot be completed or undone, a new step that leaves the whole in a safe, honest state.Confused deputyAn agent that holds a permission uses it on behalf of a caller that does not hold it, because nothing tied the action to the caller's authority.ConsensusA decision taken because several agents agree, for example two of three recommending escalate.Context idAn identifier carried by A2A messages and tasks that says which conversation they belong to, so follow-ups and refinements stay together.Context transferThe facts, history, constraints and credentials the receiving agent needs to carry on the work.ContractWhat both sides agreed on before the first message: request and result schemas, named errors, task states and versions.Coordination failureA failure in which every agent follows its own rules correctly and the group still stops, loops or amplifies.Correlation idAn id the caller puts on a request and the callee copies onto every reply, so the caller knows which question an answer belongs to.CrewAIA library that organizes agents (each with a role, goal and tools) into crews and flows, where agents hand work to coworkers through delegation tools their model calls.Custom orchestrationWriting the envelope, the transport, the task table and the routing between agents yourself, so every agreement is a line you can read and test.Dead letterA message that failed every attempt, set aside in a separate queue with the reason, instead of being dropped or retried forever.DeadlineThe point in time after which the caller stops waiting for a final answer and does something else.DeadlockTwo or more agents each waiting for another before they act, so none moves.Delegated scopeA named permission a caller passes to the next agent with its request, narrowed to what that agent needs.DelegationThe caller asks another agent to do part of the work and keeps ownership of the outcome.DiscoveryHow a caller finds an agent that can do a piece of work, and learns its address, its contract and how to authenticate.Durable queueA store between sender and receiver that keeps each message until the receiver acknowledges it.EnvelopeThe fields around a payload that say who, to whom, which exchange, which run, and with which credentials.Evidence artifactThe assessment agent's result for one submission: every piece of evidence with an id, a source and a trust mark, plus the gaps.gRPCA remote-call framework in which a service is described in Protocol Buffers and each call sends a compact binary message over HTTP/2.HandoffOwnership of the work or the conversation moves to another agent, and the first one steps out.Hierarchical topologyA tree of delegation: a top agent delegates to agents below it, which may delegate further.Idempotency keyA value the caller sends with a write so the server can recognize a repeat of the same intended write and store it only once.IdempotencyA property of an operation: repeating it leaves the system in the same state as running it once.IdentityThe name of the actor behind a request: an agent, a program or a person's role.Input-requiredAn interrupted state in which the task stops because the callee needs something from the caller, and resumes as the same task when the answer arrives.Intake agentThe agent that turns the broker's email into the submission, starts the run and talks to the broker.InteroperabilityThe property that two independently built programs can exchange messages correctly without either author changing code for the other.IsolationEverything that keeps one agent's failure, cost or compromise from spreading to others; what it can reach is its blast radius.JSON-RPCA small convention for remote calls: the request names a method and carries parameters and an id; the response carries the same id and a result or an error.LangGraphA library in which you define a shared state, nodes that read and update it, and edges that decide which node runs next; its documentation calls it an orchestration framework and runtime for long-running, stateful agents.MCP resourceA piece of context, such as a file or a document, that an application reads from an MCP server by its URI and decides whether to show to the model.MCP serverThe side of an MCP connection that provides capabilities: it lists its tools, runs them when called, and serves resources by URI.MCPA written specification for how an AI application (the host) reaches tools, resources and prompts offered by servers, using JSON-RPC 2.0 messages.Message busA small program that carries messages between the agents of the example, on a simulated clock, and can be told to misbehave.Message costEverything a message adds: latency, compute (often a model call), a place to fail, a span to store and a new holder of data.Message idAn id unique to one message, so that a second delivery of the same message can be recognized.MessageOne unit that one agent sends to another: an envelope that says who, to whom and about what, plus the payload.ModelThe language model itself: text in, predicted text out. It runs no code and keeps no memory between calls.Mutual TLSA TLS connection in which the client also presents a certificate, so each side authenticates the other before any message is sent.OAuth 2.0An open standard in which an authorization server issues access tokens that carry scopes, so a client can call a resource server without holding the user's password.Opaque agentAn agent whose tools, memory and reasoning the caller cannot see; only its messages are visible.OrchestratorAn agent or program in the middle that sends every request and receives every reply of a workflow.Out-of-order deliveryMessages arrive in a different order from the one in which they were sent.OwnershipThe agent or person responsible for the outcome of a piece of work: who must notice if it stalls and decide what happens when it fails.PartA unit of content that holds exactly one of text, raw bytes, a URL, or structured data, with an optional media type and file name.PayloadThe content of a message: what the recipient's work is about.Peer-to-peerA topology with no agent in the middle: each agent sends its request directly to the agent that does the next part.Policy fixtureFive invented underwriting rules with an id and a version, shared by every site of the series.PollingThe caller asks for a task's state at an interval until it ends.Protocol bindingThe mapping of every operation of a protocol onto one transport: the method names or URL paths, the content types and the error format.Protocol boundaryThe place in a system where traffic changes protocol because the thing on the other side changes kind: an agent that owns work, a tool that answers, or a person who decides.Push notificationThe callee notifies the caller by sending an HTTP request to an address the caller registered in advance.RecipientThe agent a message is for; the to field.RecommendationThe proposed outcome the underwriting agent writes: approve, decline, request_info or escalate, with reasons and the rule ids it rests on.RegistryA central catalog that holds Agent Cards and lets callers search them, for example by skill or tag.Request/replyOne request followed by one or more replies that carry its correlation id.Responsibility boundaryThe line between what one agent owns and what another owns: goal, tools, data, scopes and decisions.RetrySending a request or calling a tool again because the first try failed or no answer came in time.Run idAn id for one execution of the whole workflow for one submission, shared by every message and task of that execution.SchemaThe written shape of one kind of payload: fields, types, which are required. Named with a version, such as evidence.request/v1.Scripted stand-inCode that follows a fixed script where a real agent would call a language model, so the example runs without a key.SenderThe agent that wrote and sent a message; the from field.Shared stateAgents coordinate by reading and writing a common store instead of sending each other requests.Single point of failureOne component whose failure stops the whole system.SkillOne kind of work an agent advertises in its Agent Card, with an id, a name, a description, tags and examples.SpanOne timed step: a name, a start, an end, a status, and the step that caused it (its parent).SSEA way to stream over HTTP: the server answers with content type text/event-stream and writes one event at a time.State machineA list of states plus the moves allowed between them; anything else is refused.StreamingA connection the caller opens once and the callee keeps open, writing each update as it happens.SubmissionThe structured record that intake writes from a broker's email: one JSON object with the facts of one quote request.SynchronousAn exchange in which the caller waits for the reply before it does anything else.Task protocolAn asynchronous exchange in which the callee answers at once with a task id and later reports the state and result of that task.Task stateWhere a task stands at a moment: submitted, working, input-required, auth-required, completed, failed, canceled or rejected.TaskA piece of delegated work that has an id and a state, and may outlive one request.Terminal stateA state a task never leaves: completed, failed, canceled or rejected.TimeoutThe moment a waiter stops waiting for one answer and treats the step as failed, without knowing whether it happened.Tool callOne request to run one function, by name, with arguments, and receive its result; the function answers and does not decide.ToolA function the agent's program runs on the model's request. It takes arguments and returns a result; it does not decide.TopologyThe answer to one question: who may send messages to whom.Trace idThe id all spans of one trace share; 32 hexadecimal characters in the example, as in W3C Trace Context.Trace propagationPutting the trace id and the sender's span id into each message, so the receiver's span becomes a child of the sender's.TraceThe tree of all spans of one piece of work, joined by their parent links.Trust boundaryThe place where data crosses from a source you do not control into one you do, so it is checked and never obeyed as an instruction.Typed errorAn error answer with a fixed code a program can branch on, a message for a person, and whether a retry can help.Underwriter cockpitThe screen where the underwriter sees a recommendation with its reasons and records the decision.UnderwriterThe person at the insurer who decides whether to accept a risk. In this series, always a person and never an agent.Underwriting agentThe agent that reads the evidence artifact, runs the eligibility rules and writes the recommendation.VersioningGiving each shape a version and changing it when the shape changes, so both sides can say which one they speak.WebhookA URL on the caller's side that accepts notifications from the callee.Well-known URIA fixed path under a domain, defined by a standard, so a client can find a document without being told where it is.

Each term is explained in full on one page and used everywhere else. Where the words come from, and which of them this site uses as its own teaching design, is on About.