All categories/Agents and automation
Level Advanced · Tools Agent Risk
Tool contract: when not to call
Give an agent tools and it calls them constantly, with the cost and risk that implies.
Prompt
Write a tool-use contract for my agent.
Available tools:
{{tools}}
Scenario: {{scenario}}
Risk areas: {{risks}}
Format: one table, one row per tool, columns:
1. When to use it (a concrete trigger, never 'when needed')
2. When it must not be used
3. What must be confirmed before calling
4. Retry policy (how many, how long between, when to give up and escalate)
5. Where the call leaves a trace (can the user see it)
Then three cross-tool rules.What to state before it starts
Leave these unsaid and the model fills them with an assumption. Most 'wrong answers' start here.
- {{tools}}: name, one-line purpose, whether it has side effects
- {{risks}}: money, data writes, outbound messages — irreversible ones
- The trace column matters: invisible calls cause the worst incidents
Go through each of these
Where this kind of work most often goes wrong. Already in the prompt, so you do not have to remember it.
- 'When needed' is not a trigger; it voids the contract.
- A fixed retry policy stops it burning money on one failing call.