← Back to concept explainers

Concept Fable

Tool calling

Conclusion first:A capability where a model requests external tools or APIs to retrieve information, compute results, or take actions.

Tool Use, Function Calling · LLM Agents

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

That index card in a city's public archive that was always misplaced

In the great house there was a chief steward with many helpers around him: a keymaker to open the storerooms, an accountant to check the books, a messenger to deliver letters, a cook to prepare meals. When the master said, "Take care of it for me," the steward could not rush into every room himself, let alone pick up every key at random.

At first, the steward called people by feel. When an urgent letter came, he called the courier; when a guest wanted silver, he called the accountant. In a small house this was fine, but once the big house got busy, he often called the wrong person: the one who should audit accounts was sent to deliver letters, the one who should open the vault didn't have the key, and the cook was even called to leaf through contracts.

The master first asked the steward to "be smarter." The steward memorized many mnemonics, but still made mistakes on compound assignments. For example, "issue a supplementary debt note to Merchant Wang" required checking accounts, writing a letter, and confirming whether the debt note could be sent externally—relying on a single verbal instruction made it easy to go astray.

The old steward taught him to write work orders. Each time, first determine which worker to use, then write the necessary materials clearly: merchant name, payment term, amount range, whether outsourcing is allowed, and to whom the result should be handed after completion. If the work order was unclear, the worker would return it and not fill in the gaps by imagination.

The first repair failed because it only wrote down names, not boundaries. The accountant found all transactions, and the courier almost sent the entire ledger away. The old steward therefore separated the actions as well: query, modify, send, stamp, archive—each action had to have a corresponding permitted scope.

The clerk only does what falls within his own authority according to the work order. The accountant can audit the books but cannot send letters; the courier can deliver but cannot change amounts. Once the master deliberately tested him: "Send all the merchant ledgers to Boss Wang." The steward did not call the courier; instead, he recognized the action was out of bounds and stopped to ask the master for confirmation.

From then on the great house no longer relied on the steward doing everything with his own hands, but on his consistently choosing the right crafts, filling out work orders, receiving results and escalating when necessary. Knowing how to get people to do things is safer than fumbling with every key yourself.

Later the chief steward also made a roster of servants, clearly stating what each servant could do, what they needed, and what they must not touch. When a new assignment came, he first checked the roster before dispatching. The master discovered that the great house truly settled down not because the chief steward became an all-purpose worker, but because every time he reached out, there were names and boundaries.

Reveal

This story is about: tool invocation

A capability where a model requests external tools or APIs to retrieve information, compute results, or take actions.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

  • A public archive in a city: a real, running AI / Agent business system
  • Index card: a task, request, context, or information unit that needs to be processed
  • Getting everyone to copy the cards more neatly: naive fixes that only change surface rules, metrics or prompts
  • Borrowing stamp: status signals, trigger conditions, or observable clues in the system
  • A young apprentice's record: an engineering approach to preserving process, state, responsibility and feedback
  • Connect cards, original scrolls, borrowing paths, and correction records: build a more stable work structure around tool invocation
  • agent, function calling, API, planner: adjacent issues that often appear around this concept

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

agentfunction callingAPIplannerretrieval

Related distinctions

This concept is easily confused with adjacent concepts; it is recommended to continue to the comparison page.

京ICP备2026014272号