AI Agent — AI Glossary

Reference details

Related terms

• Function Calling
• Tool Use
• Agent Workflow
• MCP (Model Context Protocol)
• Agentic Loop

How it works

1. The request carries a system prompt, the tool schemas, and the task, and from that point the model's entire working state is the message list it can see.
2. Each turn the model emits either a tool call or a final answer, so the decision to continue is an output of the model rather than a branch in the harness.
3. The harness executes any call, appends the result to the message list, and returns control, which is why the transcript grows monotonically and never shrinks on its own.
4. Because every prior observation stays visible, a conclusion reached at step five can be revised at step twenty against the evidence that produced it.
5. Termination is whichever comes first: the model stops calling tools, a step limit fires, or a cost ceiling is hit — and the last two belong to the harness because the first cannot be relied on.
6. That division is the whole safety story of the architecture: the model chooses the path, and every bound on how long and how expensively it may wander is imposed from outside.

Example: A chain vs. an agent, on the same task

Task: “What did we bill Acme last quarter, and is that above their contract cap?” A chain runs fixed steps: query invoices → sum → compare → answer. If the invoice table is empty it returns £0 and stops. An agent decides its next step from what it sees: the empty result prompts it to check whether Acme is billed under a parent account, re-query, then compare. Same tools, same model — the difference is who chooses step two.

Commonly confused with

Not every LLM app that calls a tool is an agent. A single function call inside a fixed pipeline is tool use in a chain. The distinguishing property is control: an agent selects its next action at runtime, so its execution path is not knowable in advance.

When to use it

Reach for it when:

• The number of steps depends on what earlier steps return
• The task needs recovery from partial or failed results
• You can define a stopping condition and a step budget

Reach for something else when:

• The steps are fixed — a chain is cheaper, faster and far easier to test
• Every action is irreversible and unattended, with no human review
• You cannot bound cost: a loop that picks its own next step can pick badly many times

Referenced by

This term is used by the following ModelRefs references:

• tree of thoughts
• function calling prompt
• react
• plan and execute
• reflexion
• react
• plan and execute
• single tool agent
• multi tool router
• code execution agent
• short term memory
• long term memory rag
• supervisor worker
• debate
• chain of agents
• reflexion

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to AI Agent — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path