All chapters
  1. Start hereThe model proposes a call. Your code decides whether it runs
  2. Foundations · 1Functions: arguments, return values and side effects
  3. Foundations · 2Tool calling: how a model asks, and the loop that answers
  4. Foundations · 3Tool schemas: names, descriptions and argument constraints
  5. Foundations · 4Types and annotations: one source for the schema and the handler
  6. Foundations · 5Result contracts: success, error, partial result, reference
  7. Protocols · 6Discovery and selection: how the right tool gets chosen
  8. Protocols · 7MCP architecture: host, client, server, tools, resources, prompts
  9. Protocols · 8MCP on the wire: transports, lifecycle, authorization, notifications
  10. Protocols · 9Retrieval: a tool the model calls, or knowledge it is given
  11. Production · 10Implementing tools: API wrappers, database queries, files, code execution
  12. Production · 11Safety: least privilege, validation, idempotency, sandboxing
  13. Production · 12Reliability: timeouts, retries, rate limits, circuit breakers
  14. Production · 13Testing tools: contracts, stand-ins, faults and traces
  15. Reference · 14Frameworks and standards: what each one does for a tool
  16. Reference · 15Advanced: composition, pagination, streaming, binary content, observability
  17. Reference · 16The insurance lab: one submission, five tools, end to end
  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, what it is, what it is not, and how it shows up in the insurance example.

Two sentences hold the words apart. The model proposes a call; your code decides whether it runs. And: the model recommends; a person decides. A tool call is a request written by the model. Execution is what the application does after it has validated and authorized that request.

Agent loopThe loop in which the application calls the model, executes the tool calls it returns, appends the results, and calls it again until it stops asking.AgentA program that calls a language model again and again, running the tools the model asks for in between, until the model gives a final answer.API wrapperA function that calls another system's interface and translates its answers and failures into the tool's own terms.ArgumentA value the caller passes into a function, telling it what to work on.Attempt numberA count of the tries of one tool call inside a run: 1 for the first, 2 for the first retry, 0 when nothing ran.BackoffA wait before each retry that grows with every failure, usually doubling, with jitter added so callers do not retry in step.BackpressureMaking overload visible to whoever sends work, so that work waits or is refused at the entrance instead of piling up in the middle.Binary contentData that is not text, such as an image, a PDF or a sound file, returned by a tool.Capability negotiationThe way an MCP client and server declare which features they support, so that neither uses a feature the other lacks.ChunkA piece of a larger document, a few sentences long, stored and retrieved as one unit with its own labels.Circuit breakerA small state machine per provider that, after repeated failures, makes calls fail at once without calling the provider, and lets one trial call through after a cool-down.CitationA reference from a statement in a recommendation to the rule or evidence it rests on, so a person can check it.Code execution toolA tool that runs code the model wrote, inside an isolated environment, and returns its output.Confused deputyA program that holds more authority than the party asking it to act, and is tricked into using that authority on that party's behalf.Contract testA test that checks the shape of an interface (fields, types, allowed values) rather than one behavior behind it.DeadlineA fixed point in time, set when a call starts, after which no attempt starts or runs.DeserializationThe step that reads a JSON object and builds a typed value from it, where wrong types, missing fields and unknown fields are found.DTOA type with fields and no behavior, used to carry data from one system to another, such as a request body or a tool's arguments.EmbeddingA list of numbers computed from a piece of text so that texts with similar meaning get nearby lists, which lets a search find passages by meaning rather than exact words.EnumA JSON Schema keyword that restricts a value to one of a fixed list.Error resultA tool result that reports a failure as readable data, with a code, a message and a retry hint, instead of an exception the model cannot catch.Evidence idThe identifier of one piece of evidence a tool returned, cited by the recommendation so that a person can open exactly what the agent relied on.Execution adapterThe code path that looks a requested tool up, validates its arguments, authorizes the caller, runs the function with limits, and shapes the outcome into a result.FastMCPAn open-source Python framework for building MCP servers and clients, separate from the official SDK.Fault injectionMaking a dependency fail in a chosen way at a chosen moment, so the failure path runs under test.Function callingThe mechanism by which a model, given tool definitions, answers with structured requests for tools instead of, or as well as, text.FunctionA named piece of code that takes values in, may change something, and gives a value back.Human approvalA control under which a write runs only after a person has seen the exact call and said yes.Idempotency keyA value chosen by the application, the same for every attempt of one logical write, that lets the store return the earlier result instead of writing again.IdempotencyThe property that repeating an operation leaves the world as one execution would have left it.Input schemaThe part of a tool definition that says, in JSON Schema, which argument objects are valid.Integration testA test that runs several real parts together across their real boundaries, so mistakes between the parts show up.Java recordA Java class declared by a list of components, with final fields and an automatically derived canonical constructor and accessors; delivered in Java 16.JSON SchemaA standard way to describe which JSON values are valid, using keywords such as type, properties, required, enum, pattern and additionalProperties.JSON-RPCA small standard for remote procedure calls written as JSON: a request has a method, params and an id; the response carries the same id and a result or an error.LangChain toolA Python function marked with LangChain's @tool decorator, which takes the name from the function, the description from the docstring and the schema from type hints.LangChain4jA Java library for model applications in which @Tool methods become tools for its AI Services.Least privilegeThe rule that a caller is given only the permissions its task needs, so a mistake or a steered model can do only that much harm.MCP authorizationThe part of MCP that says how a remote server learns who is calling and what they may do, using OAuth 2.1.MCP clientA connector the host creates for exactly one server: it sends requests with the protocol version and capabilities attached, and passes results back to the host.MCP hostThe application that holds the conversation, talks to the model, creates one MCP client per server, and enforces consent and security policy.MCP lifecycleHow an MCP exchange starts, runs and ends; in revision 2026-07-28 every request stands alone.MCP promptA message template a server offers for a user to select explicitly, often as a slash command, filled with arguments the user gives.MCP resourceData a server offers under a URI for the host application to bring into context, by its own choice rather than the model's.MCP SDKA library that implements MCP so that you write tools as functions and it writes the protocol messages.MCP serverA program, local or remote, that offers tools, resources and prompts through MCP and executes the tool calls it receives.MCPAn open protocol through which a server offers tools, resources and prompts, and any host application that speaks it can discover and use them.MockA test replacement for a dependency that records the calls it receives so the test can assert on them.NamespaceA prefix on a tool's name that names its service or domain, so the model knows the domain at a glance and two teams' tools cannot collide.NotificationA JSON-RPC message without an id; the receiver must not reply, so the sender gets no confirmation.OAuthA standard way for a program to get a token that lets it call a service on someone's behalf without seeing their password.OverrideA person's decision that differs from the agent's recommendation, recorded as such.PaginationSplitting a long result into pages, with a cursor that says where the next page starts.Parallel tool callsSeveral tool calls written in one model turn, each with its own id.Parameter validationChecking, in the application's own code, that every argument a model wrote matches the tool's input schema before the function runs.Parameterized queryA database statement written with placeholders, whose values are passed separately so that a value can never change what the statement does.Partial resultA successful result that holds only part of the answer, marked partial, with a cursor the model can pass back to get the next part.Path traversalAn attack or accident in which a file path such as ../../etc/passwd climbs out of the folder a tool was meant to read from.Policy fixtureThe insurer's written underwriting guidance as versioned data: rules with ids, grouped in versions. Illustrative, not insurance practice.Progressive disclosureShowing the model a small set of tools at first and revealing more only when the task needs them, instead of sending the whole catalog every time.Prompt injectionText that reaches the model as data, in a tool result, an email or a document, and tries to make the model follow it as an instruction.Provider API and SDKThe interface, and the client library around it, that one model vendor offers for sending requests with tool definitions and receiving tool calls.Pure functionA function whose result depends only on its arguments and which changes nothing outside itself.PydanticA Python library whose model classes declare a shape with type hints and Field constraints, validate data against it, and generate a JSON Schema from it.RAGFetching relevant text from a store and putting it in the model's context before it answers, so the answer can rest on documents the model was not trained on.Rate limitA cap a provider puts on how many requests a caller may send in a period; over it, requests are refused, usually with a time to wait.Read toolA tool whose function changes nothing outside the run; calling it again is harmless apart from its cost.RecommendationOne of approve, decline, request_info or escalate, with reasons, cited rule ids, anything missing, and requires_human true.Required fieldA property listed in the required array of an object schema; leaving it out makes the object invalid.Resource referenceA part of a result that names where larger content can be read (a uri, its media type and size) instead of including it.Result envelopeA wrapper around every tool result: ok, then either data or an error (code, message, retryable), plus meta about the call. This site's design, not a standard.Retrieval toolA tool whose job is to search a store and return passages, where the model decides when to call it and writes the query.RetryA second or later attempt at the same tool call after an attempt failed in a way that may pass.Retryable errorAn error result whose flag says that asking again, now, may help, because the cause was slowness, load or a passing fault rather than the request itself.Return valueThe value a function gives back to its caller when it finishes without failing.Roof-data providerThe outside service that assesses a building's roof and returns its age; a stand-in in the example.Run idThe identifier of one execution of the workflow for one submission.SandboxAn isolated environment in which code runs with only explicitly granted access: no network unless allowed, a limited file system, no credentials, and limits on memory, CPU and time.Schema generationTurning a function's type hints, a model class or a Java type into the JSON Schema a tool definition carries, so the schema cannot drift from the code.ScopeA label for one kind of access, such as uw.read or uw.recommend, that a caller either holds or does not.SecretA value such as an API key, a password or a token that lets its holder use a system, and so must never reach the model.SessionState that a server keeps between requests of one client; MCP revision 2026-07-28 removed protocol-level sessions.Side effectAny change a function makes outside its own return value: a row stored, a file written, a message sent.SpanA named piece of work with a start, a duration, a status and attributes, recorded as part of a trace.Spring AISpring's framework for model applications, in which a method annotated @Tool becomes a tool.Stand-inA scripted replacement that plays the part of the model or a provider, so a run gives the same result every time.stdio transportAn MCP transport in which the client starts the server as a subprocess and exchanges one JSON-RPC message per line through its standard input and output.Stop reasonThe recorded reason a turn or a run ended, such as a request for tools, a final answer, or a limit.Streamable HTTPAn MCP transport in which every client message is an HTTP POST to one endpoint, answered with JSON or an event stream for that request.Streaming resultA result that arrives over time, as progress updates or parts, instead of in one message.Strict modeA switch in a provider API that makes the model's tool arguments follow the input schema while they are generated, for a documented subset of JSON Schema.Structured contentThe structuredContent field of an MCP tools/call result: any JSON value, which must conform to the tool's outputSchema when one is declared.SubmissionOne JSON record that holds the facts of a broker's quote request; every step after intake reads it instead of the email.TimeoutThe longest the application waits for one attempt before treating it as failed.Tool annotationMetadata on a tool definition that tells the client how the tool behaves: whether it only reads, may destroy, is safe to repeat, or reaches outside systems.Tool call idThe id the model, or the provider for it, gives one tool call; the result goes back under the same id.Tool callThe structured request the model writes: a tool name, the arguments, and a call id. It is a proposal, not an action.Tool choiceA request option that lets the model decide, forces it to call some tool or one named tool, or forbids tool calls.Tool compositionUsing the result of one tool call as input to another call.Tool definitionThe object the application sends with a request to describe one tool: a name, a description and an input schema.Tool descriptionThe sentences in a tool definition that say what the tool does, when to use it and when not, whether it reads or writes, and what comes back.Tool registryThe application's complete list of the tools it can offer, with everything needed to run each one: definition, function, required scope, provider and timeout.Tool resultThe message the application sends back to the model for one tool call, keyed by that call's id.Tool searchA search tool the model can call to find and load tool definitions that were sent with the request but kept out of its context.Tool selectionThe model's decision, made from the request and the tool definitions alone, about which tool to call next, if any.Tool use blockThe content block in which Anthropic's Messages API returns a tool call: type tool_use, with id, name and input.Tool versioningChanging a tool's definition or behavior in a controlled way so that callers that depend on the old one keep working.ToolAn external capability the agent may request: the model proposes a call, and application code validates and executes it.Trace checkAn assertion over the recorded trace of a run (which tools were requested, in which order, with which outcome) rather than over its final answer.TraceThe record of one run as a tree of timed steps, identified by one trace id.TransportThe way messages travel between a client and a server, such as a pipe between processes or HTTP.Type hintAn annotation on a Python parameter or return value, such as submission_id: str, read by type checkers and libraries but not enforced when the code runs.Underwriter cockpitThe web app in which an underwriter reads the agent's recommendation with its reasons and evidence, and records the decision.UnderwriterThe employee of the insurer who decides whether to accept a risk. In the example, the only party who decides.Underwriting agentThe agent that reads a submission, gathers evidence with five tools, applies the rules and writes a recommendation for a person.Untrusted contentAny text that was not written by the application or the person it serves, such as retrieved passages, emails, attachments or other servers' results, read as data and never followed as instructions.Vector storeA store that keeps embeddings with their passages and labels, and answers a query embedding with the passages whose embeddings are nearest.Write toolA tool whose function changes something outside the run, so a repeated call repeats the change unless code prevents it.

Each term is explained in full on one page and used everywhere else. Statements about named products, protocols and libraries were read in their own documentation on 2026-11-17; the sources are listed on About.