← Back to concept explainers

Concept Fable

Tool selection

Conclusion first:Tool Selection is the process by which an Agent, based on the current subtask, chooses which available tool, function, or API to call and with what parameters. It is not just capability matching; it also involves intent recognition, permission boundaries, risk level, cost, reversibility, and whether human approval is needed. In enterprise Agents, tool selection determines whether the system can reliably execute real actions, avoiding wrong tools, excessive authorization, parameter errors, or turning low-risk queries into high-risk write operations.

Tool Selection · Agent decision-making

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

Fable

Once there are many keys, opening the wrong door is more dangerous than having no key at all

The service hall of Sea Salt City had a row of keys. The red key opened the archive cabinet, the bronze key opened the accounting room door, the black key opened the warehouse, and the silver key could issue announcements. In the past there were few keys, and an old clerk could tell at a glance which one to take for a request.

Later the hall connected to more places. You could check tax ledgers, send letters, change inventory, book carriages, issue invoices, freeze accounts. Every key was useful, and every key could cause trouble. What newcomers feared most was not having no keys, but not knowing whether they should use a given key.

The steward first posted a big notice: use the red key for archive issues and the bronze key for accounting issues. The notice quickly became useless. A merchant supplementing an invoice involved both accounting and archives; a warehouse return required changing inventory and might also trigger a refund. Real tasks never line up according to the boxes on the notice.

Someone suggested giving all the keys to the cleverest clerk and letting him choose by feel. For the first few days it was fast, until he opened the complete accounting ledger just to look up a customer address; and to confirm inventory, he modified an order that had not yet been approved. The work got done, but the risk came in with it.

The worst time was when the silver key was used to post a public notice. The clerk meant to send a draft to one merchant, but posted an unreviewed price list at the city gate. It wasn't that he wasn't smart—it was that he wasn't required to first assess action risk, parameter boundaries, and whether approval was needed.

Later the hall changed key selection into three steps. First look at task intent: is it to query, to modify, to notify, or to create a new record. Then look at tool capabilities and limits: which key can do this thing, which can only read and not write. Finally look at risk: whether it involves money, identity, external notification, or irreversible operations.

Each time the clerk takes a key, he must write the reason for the choice and the key parameters on a small tag. Low-risk queries can be done directly; changing accounts, issuing announcements, and deleting records must be stamped by the manager; if either key can do the job, choose the one with narrower permissions and more reversible results.

The hall did not slow down because of this extra set of choices. On the contrary, rework and incidents decreased. Everyone understood that the real skill is not having the most keys, but choosing the one that is strong enough for each subtask without over-authorizing it.

Reveal

This story is about: tool selection

Tool Selection is the process by which an Agent, based on the current subtask, chooses which available tool, function, or API to call and with what parameters. It is not just capability matching; it also involves intent recognition, permission boundaries, risk level, cost, reversibility, and whether human approval is needed. In enterprise Agents, tool selection determines whether the system can reliably execute real actions, avoiding wrong tools, excessive authorization, parameter errors, or turning low-risk queries into high-risk write operations.

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 row of keys: the tools, functions and APIs available to an Agent
  • The big notice: simple keyword rules struggle to cover real business combination scenarios
  • The clever clerk chooses by feel: relying solely on the model's free judgment, lacking policy constraints
  • Silver key mistakenly issued in an announcement: wrong tool selection caused an externally visible incident
  • Three steps—intent, capability, risk: the core judgment dimensions for tool selection
  • Selection rationale and key parameters: auditable tool selection trace
  • Manager's stamp: human approval for high-risk tool calls
  • A narrower key: enterprise execution principles prioritizing least privilege and rollback capability

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.

tool usefunction callingrouterexecutorReAct

Related distinctions

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

京ICP备2026014272号