Glossary

Terms used across the agentic workflows lessons, from the agent loop to multi-agent design.

A

Agent

A system in which the model chooses its own next step and tool use in a loop, rather than following a path your code fixes.

Agent loop

The repeating cycle of calling the model, running any requested tool, returning the result, and calling the model again until it finishes.

Agent-computer interface (ACI)

The design of tools, parameters, and return values as seen by the model. Anthropic recommends investing in it as much as in a human-facing interface.

Agentic system

The umbrella term Anthropic uses for both workflows and agents. Use it when you do not need to specify which one.

Attention budget

Anthropic's term for the finite capacity of a model to draw on the tokens in its context. Each added token spends some of it.

Augmented LLM

The basic building block of agentic systems: a model enhanced with retrieval, tools, and memory.

C

Capability eval

An eval that measures what an agent can do now. It is expected to pass only part of the time at first, and shows progress.

Code-based grader

A script that checks an agent's output, such as a test suite or a query result comparison. Cheap, repeatable, and preferred wherever the answer is checkable.

Compaction

Summarizing a long conversation and starting a fresh context window from the summary.

Context rot

The decline in a model's ability to recall information accurately as the number of tokens in its context grows.

Context window

The set of tokens the model sees when it generates a response. In an agent it includes instructions, tool definitions, messages, and tool results.

E

end_turn

The stop reason that means the model has finished its answer and has no further tool requests.

Eval

A test that gives an agent an input and applies grading logic to what it produces.

Evaluator-optimizer

A workflow pattern in which one model call generates output, another critiques it, and the loop repeats against clear criteria.

G

Guardrail

A control in your code that limits what an agent may do or how long it may run. Budgets, permissions, approvals, and validation are guardrails.

H

Human grader

A person who reviews an output. Used to calibrate model-based graders and to spot-check results.

Human-in-the-loop approval

A gate that pauses a run until a person approves or rejects a specific tool call.

J

Just-in-time retrieval

Giving an agent references such as file paths or queries, and letting it load the data only when needed.

M

Model-based grader

A model that scores an output against a written rubric. Use it where code cannot judge quality, and check it against human review.

O

Orchestrator-workers

A workflow pattern in which a central model decides the subtasks at runtime and delegates them to workers.

P

Parallelization

A workflow pattern that runs independent subtasks at once (sectioning) or runs the same task several times (voting).

pass@k

A metric asking whether at least one of k attempts succeeds. Use it when one success is enough.

pass^k

A metric asking whether all k attempts succeed. It is the stricter measure and usually the right one for customer-facing agents.

Poka-yoke

A design practice that makes errors hard or impossible to make. In tool design, it means changing arguments so a wrong value cannot be passed.

Prompt chaining

A workflow pattern in which sequential model calls each process the previous output, with optional programmatic gates between steps.

R

Regression eval

An eval that protects behavior that already works. It should stay near full pass, and confirms that a change did not break anything.

Routing

A workflow pattern that classifies an input and sends it to a specialized path, such as a smaller model for easy cases.

S

stop_reason

The field in a model response that tells your loop why the model stopped. The two values used in this course are tool_use and end_turn.

Step limit

A maximum number of loop iterations enforced by your code. It is a budget guardrail that the model cannot override.

Structured note-taking

The agent writes progress notes to storage outside the context window and rereads them later.

Sub-agent

An agent that works in its own context window on a focused task and returns a short summary to a lead agent.

T

Tool definition

The name, description, parameter schema, and return contract that tell the model what a tool does and how to call it.

Tool result clearing

Removing old raw tool outputs from the context once the agent no longer needs them. The lightest form of compaction.

tool_result

The message your code sends back after running a requested tool, containing the output or an error the model can act on.

tool_use

The stop reason, and the content block, in which the model requests that your application run a tool.

W

Workflow

A system in which LLMs and tools follow predefined code paths that your code controls.

No matching terms. Try a different word.