Every term this site uses, in one place. Each card opens the same drill-down as the underlined phrases in the text: one line, why it matters here, what it is, what it is not, and how it shows up in the running example.
Abstraction ladderThe ordered list of layers you can stand on to run a model call: each wraps the one below and hides a named piece of work.Agent frameworkA library that runs the loop and ships ready-made versions of several jobs around it, usually for many model providers.Agent loopThe code that calls a model again and again, running the tools it asks for in between, until the model stops asking.Agent SDKA library that runs the agent loop for you: call the model, run the tools it asked for, send the results back, repeat.AgentA model call in a loop: the model asks for a tool, code runs it and sends the result back, until the model stops asking.Allow-listThe harness runs a tool only if its name is on the list. Everything else is refused.Approval gateA point where the harness stops and waits for a person before it runs an action that cannot be taken back.CheckpointA copy of everything the run needs to continue, written outside the process after each step.ChunkA piece of a document, stored and embedded on its own, so that a search can return the part that matters and not the whole.Coding harnessA ready-made agent harness for changing software: it runs the loop, ships file and shell tools, and asks the developer before risky steps.CompactionOld turns of a run are replaced by a short summary, so the run can continue inside the context window.Context assemblyThe harness job of deciding what goes into each request to the model, and in what order.Context windowThe size limit, in tokens, on one call's input plus its output.Dead-letter queueA queue for events that failed every retry, kept so a person can inspect and replay them.Durable executionThe progress of a run is recorded outside the process, so another process can continue it after a crash, a deploy or a long wait.Ecosystem mapA table with one row per product and one column per job, where each cell says who runs that job: you, the product, a named partner, or the vendor as a service.Eight jobsThe eight pieces of ordinary software every agent system needs around the model call, whoever runs them.EmbeddingA fixed-length vector that an embedding model returns for a text, so that texts with similar meaning get nearby vectors.Eval harnessThe rig around an eval: load the cases, run each through the agent, grade, add up, report.EvalKnown inputs run through the agent; a grader compares the result with what was expected.Fake modelCode that takes the same request as the model endpoint and returns the same response shape, but decides with fixed rules.Fallback modelWhen waiting does not help, the harness sends the same request to a different model.Golden setA kept list of inputs, each with the result a person has confirmed is correct.Graph runtimeA runtime where you describe work as nodes and edges, and it saves the state as the run proceeds, so a run can pause and resume.GuardrailA check that the harness enforces around the model. It holds whether or not the model follows its instructions.HarnessThe software around the loop: it builds the context, runs the tools, keeps the state, enforces the rules and records what happened.HookA script or other handler that the harness runs at a named point, for example before every tool call. The harness runs it; the model does not.Human in the loopAt a chosen point the harness stops, saves the run, and continues only when a person has answered.Hybrid searchRunning a keyword search for exact tokens and a vector search for meaning, then combining the two rankings into one list.IdempotencyAn operation is idempotent when repeating it changes nothing more. A key makes a side effect idempotent.Instructions fileA text file, usually in the repository, that a coding harness adds to the context of every session.Judge modelA second model call that reads an output and a rubric and returns a grade.Least privilegeGive a role, a stage or a tool only the tools and data its task needs, so a mistake can reach only that much.Managed agentAn agent you define and a vendor runs: you send a definition and messages, the vendor's service runs the loop and stores the session.Managed harnessA service in which a vendor runs the agent loop on its own servers, usually with the tool sandbox and the session state.MCP clientThe component inside the host application that sends MCP requests to exactly one server and returns the replies.MCP serverA separate program that answers MCP requests: it lists what it offers and runs a tool when a client calls one.MCPAn open protocol with which a separate program offers tools and data, and any harness that speaks it can list them and call them.MemoryFacts stored outside the context, by the harness, and loaded into a later run's context when they are needed.Messages APIOne HTTP request in, one JSON response out: POST /v1/messages.MiddlewareA function that a framework's loop calls around a model call or a tool call, so you can inspect, change, retry or stop that call without editing the loop.ModelThe language model itself: it receives one request and returns text or a tool request. It keeps nothing between two calls.OrchestratorThe part of a system that decides which agent runs next and what it is given. It can be code, or an agent whose model asks for other agents through tools.OutboxWrite the intended effect to your own database in the same transaction as the state change; a worker delivers it.Output contractThe agreed structure of the final answer, checked by code before another program reads it.Permission modeA named default for the approval gate of a coding harness, from asking about every change to asking about nothing.Prompt cachingThe provider stores the processed start of a request, so the next request with the same start is cheaper and faster.Prompt injectionText from an outside source that tries to make the model follow it as if it were an instruction.Provider SDKA client library from the model vendor that wraps the HTTP call: headers, retries, streaming and typed objects.RAGCode finds the records a request needs and puts them into the context, and the model answers from them.Re-rankingA second stage that takes the candidates from a fast, wide search and orders them again with a slower, more precise rule.Retry with backoffAfter a failure that may pass, wait and try again; double the wait after each failure.SandboxLimits on files, network and resources that hold whatever the code tries to do.SkillA folder with a file of instructions, usually SKILL.md. Only its short description is in the context until the skill is used.StatelessEach call is answered from its own contents. The next call starts from zero.Stop conditionA check the harness makes after each step. When it is true the loop ends with a named status.Stop reasonA field in every response that tells the calling program what state the reply is in.StreamingA way to receive a model's reply as a series of small events on one open connection, instead of one complete body at the end.SubagentA second run of the agent loop, with its own instructions, tools and context, started by a harness for one task. Usually only its final answer goes back.System promptText in the request that tells the model its role and rules, apart from the conversation.Token budgetThe harness adds up the usage of every response and stops the run when the total reaches a limit.TokenA small piece of text: a word, part of a word, a character or a byte.Tool requestA block in the model's reply that names a tool and gives arguments for it. It is a request, not an action.Tool resultA block the harness sends in the next call, carrying a tool's output and the id of the request it answers.Tool runnerA beta helper inside Anthropic's client SDKs that calls the model, runs the tool functions it asks for, returns the results and repeats.Tool schemaThe part of a tool definition that says which arguments exist, their types, and which are required.Tool serverA program that owns a set of tools, their credentials and their rules, and serves them over a connection.ToolA function with a name, a description and an input schema. The model asks for it; the harness runs it.TraceOne entry per model response and per tool call of a single run, all carrying the same run id.TransportThe channel that carries the JSON messages. The meaning of the messages is the same on every transport.TurnOne pass through the agent loop: one model call, then the tools that reply asked for.Vector storeStorage that holds one vector per document and answers the question: which stored vectors are nearest to this one?WorkflowA sequence of steps whose order is fixed by code. A model, if one is used, works inside a step and never chooses the next step.