Agents: a step that decides which tools to use
When the right sequence depends on the item, give a model an allowlist of your tools and let it work — read-only by default, capped, and with a full transcript afterwards.
In this article (4)
A Call a tool step calls one specific tool every time. An Agent step is for the other case: the answer needs two lookups for this item and four for that one, and you would rather not encode every branch by hand.
Setting one up
Add an Agent card and pick the tools it may use from your connected servers. It then runs a bounded loop per item: think, call a tool, read the result, continue, answer.
The agent must run on a Claude (Anthropic) pool — that is the adapter that forwards tool calls (and images) properly.
The guardrails, all of them tested
- Allowlist only. Nothing outside the tools you ticked is available to it, ever.
- Read-only by default. A tool that can change things requires a second, explicit "allow changes" tick.
- Capped: at most 5 tool calls and 4 turns per item by default. Reaching a cap ends the step cleanly with what it has, rather than spinning.
- A tool that reports an error is not retried. It might have had an effect; blind retries are how automation does damage.
- After a side effect, retries are downgraded for that step for the same reason.
- Full transcript. Open Steps on an item to see every call, its arguments and its result. An agent you cannot audit is not one to trust with a catalogue.
Saved agents and skills
- Settings → Agents stores a reusable agent: a system instruction ("training"), a knowledge or rubric block, and a default provider and model. Any Agent card can then just pick it — so your QA scorer behaves identically in every flow that uses it.
- Settings → Skills stores named instruction packages (SKILL.md-style markdown) that you can attach to an agent or drop in as a step-instruction preset. Paste the content or fetch it from a URL.
Agent or Call a tool?
| Situation | Reach for |
|---|---|
| Always the same tool with the same shape of arguments | Call a tool — cheaper, predictable |
| "Answer this using whatever data you need" | Agent |
| A verdict on an image (QA scoring) | Agent or a plain AI step on a vision model |
Cost note: an agent makes several model calls per item — one per turn — so it costs more than a single AI step. The caps bound that, and every turn is booked and visible per item in the run monitor.