Give an Agent business context without excess authority

Use this when an Agent needs internal vocabulary, output rules, and business tools. The result is hot-reloadable context, explicit tool authority, risk approval, and replayable runs.

Four independent control planes

Plane	Owns	Does not own
Prompt Section	Vocabulary, rules, and tool guidance	Tool or data authority
Tool Feature	Input schema, execution, and model output	Host execution policy
requiresApproval	Which calls require a person	Tool visibility
ai.agent	Execution preset, allowlist, and iteration limit	Prompt content

Implementation

1. Declare required or optional sections under prompt-sections/ with explicit budgets.
2. Expose minimal tools under tools/ with structured schemas. Side-effecting tools must not bypass approval by default.
3. Start with execSecurity: deny or allowlist; permit only required commands in known working directories.
4. In Console Runtime Capabilities, inspect Prompt Section owner, source, profile, and budget policy.
5. Run one read-only and one side-effecting task in Agent Studio. Verify approval, cancellation, trace, and artifacts.

Acceptance

• A failed reload cannot change the fixed Prompt Section snapshot of an active turn.
• Console proves that a section entered the current generation without exposing content or metadata.
• A cancelled tool is not projected as a normal completion; side effects remain auditable.
• Working directory and security policy are explicit run inputs, never elevated from chat text.

See Agent tools and Prompt Sections and Agent deep dive.