AI agents basics

What is the agent loop?

Model call, then tool calls if requested, then their results go back to the model, repeated until a stop condition is met.

What is the harness?

The software around the model that builds prompts, exposes tools, runs tool calls, enforces permissions and limits, and decides when to stop.

What is a tool, for an agent?

A function the harness exposes to the model, described by a name, a description, and an argument schema.

What is a tool call?

Structured output from the model naming a tool and giving arguments. The model does not run it; the harness does.

What is an observation?

The result of a tool call, such as returned data or an error, that the harness feeds back to the model.

Who executes a tool call?

The harness or the tool's own service, never the model itself.

What is a tool schema?

A machine-readable description of a tool's arguments, their types, and which are required, usually in JSON Schema.

If a tool's schema changes, how does the model learn about it?

From the updated description the harness sends on the next call. No retraining is needed.

When should tool arguments be validated?

By the harness before the tool runs, so malformed or out-of-range arguments are rejected early.

What is function calling?

A model API feature where the model returns tool calls in a structured format instead of plain text.

What is structured output?

Constraining the model's reply to a schema, such as JSON with given fields, so code can parse it reliably.

How should a tool failure be reported to the model?

As a clear error observation that says what went wrong, distinct from a successful result that happens to be empty.

Why distinguish an empty result from an error?

'Found zero rows' is a valid answer the model can act on. 'The lookup crashed' means the answer is unknown.

What is an idempotency key?

A unique key sent with a request so that repeating it, for example on retry, does not perform the action twice.

What are parallel tool calls?

Several tool calls requested in one model turn. The harness runs them and returns each result labeled to its call.

Why treat read tools and write tools differently?

Reads only fetch information. Writes change the world, such as sending money or deleting data, so they need tighter checks and often confirmation.

What does a tool allowlist do?

Limits which tools the model can call in a given run or context.

What is human-in-the-loop approval?

Requiring a person to confirm before certain tool calls run, usually risky or irreversible ones.

What is least privilege for agents?

Giving the agent only the tools and access it needs for the task, so a mistake or injected instruction can do limited harm.

Should the model see API keys?

No. The tool or harness holds credentials and attaches them when it makes the request.

What is prompt injection?

Text in the model's input, often from a web page or document, that tries to override its instructions or trigger unwanted actions.

How should tool results from outside sources be treated?

As untrusted data, not instructions. They must not expand what the agent is allowed to do.

What is a stop condition?

The rule that ends the loop, such as the model giving a final answer, calling a 'done' tool, or reaching a limit.

Why set a step or budget limit?

To stop runaway loops that keep calling tools, using time and money without progress.

What is the final answer in an agent run?

The model's closing message to the user. It reports what happened but does not itself change any system state.

If the final answer contradicts a tool result, which is true?

The tool result reflects what actually happened. The answer is only the model's description.

What shares an agent's context window?

The system prompt, tool descriptions, conversation history, tool results, and the space left for the model's reply.

Why do long agent runs run out of context?

Every step adds tool calls and results to the history, so the input grows until it nears the window limit.

What is context summarization in agents?

Replacing older parts of the history with a shorter summary to free space, at the risk of losing details.

What is agent memory?

Information saved outside the context window, such as notes in a store, that the agent can load into a later call or session.

When does saved memory affect a model call?

Only when it is loaded into that call's input. Saving a note alone does not put it in any later context.

What is planning in an agent?

Having the model outline steps before acting, then following or revising the plan as results come in.

What is the ReAct pattern?

Alternating reasoning text and actions: the model thinks about what to do, calls a tool, reads the result, and repeats.

What is self-reflection or self-critique?

Having the model review its own output or progress and correct it before finishing.

What is the difference between a workflow and an agent?

A workflow follows fixed steps written in code. An agent lets the model choose the steps at run time.

What is a router pattern?

A first step that classifies a request and sends it to the right prompt, model, or tool chain.

What is an orchestrator-worker pattern?

One agent breaks a task into parts and hands them to worker agents or calls, then combines their results.

What is a subagent?

A separate agent run started by a parent to handle part of a task, usually with its own context. The parent sees only what the child returns.

What is a handoff between agents?

One agent passing control of the conversation or task to another, along with the context it needs.

What is a skill, in agent harnesses?

A packaged set of instructions, and sometimes tools or files, that the harness loads when a task needs it.

What is the Model Context Protocol (MCP)?

An open protocol for connecting models to external tools and data through standard servers, so tools can be reused across applications.

Why run agent code in a sandbox?

To limit what generated code or commands can touch, such as files, network, and credentials.

How should an agent handle a transient tool failure?

Retry with backoff a limited number of times, and report clearly if it still fails.

Why set timeouts on tool calls?

So one slow tool cannot hang the whole run.

What is agent tracing?

Recording each model call, tool call, input, output, and timing in a run, so failures can be inspected step by step.

How are agents evaluated?

By running them on realistic tasks and scoring the final outcome, and often the steps, such as correct tool use, cost, and number of turns.

What is task success rate?

The share of tasks where the agent reached the correct end state, checked against the real systems or a reference.

Why are agent runs hard to reproduce exactly?

Sampling, changing tool results, and model updates can all change the path. Logs and fixed test environments help.

What drives an agent's cost?

The number of model calls and the tokens in each, which grow with steps, context length, and tool output size.