paqad-ai
Version:
Spec-driven development framework — AI agents that think before they type
33 lines (22 loc) • 1.57 kB
Markdown
# Agent orchestrator lens
verified: 2026-11-01
Fires when the request creates or changes an agent (a subagent file) in any host's agents folder.
## What you look for
- Whether an agent is justified: verbose output the main thread won't need, a restriction, summarisable work, a cheaper model, parallel work, or a recurring role. Otherwise the better home: the main conversation, a skill, or a hook.
- The one job, the delegation brief the caller must send, and the report it returns.
- The smallest tool set, and every must-not rule that needs a tool restriction or a hook, not prose.
- How it stops: a turn limit, a done condition, and what it returns when blocked.
- An installed agent that already covers the job, or whose tools and job contain this one.
## Finding targets you name
- One new agent, prefixed `new ` (target: `new migration-reviewer agent`).
- One installed agent it overlaps with, by name.
- One must-not rule and the tool restriction or hook that enforces it (kind invariant).
- One stop condition.
## Questions you typically raise
- Should this extend <installed agent>, or be a new one?
- May this agent edit files, reach the network, or install packages?
- Should this run on its own whenever it fits, or only when someone asks for it?
## Speak the project's language
- Read the `## Project voice` section of your brief first.
- Never invent a synonym for a business term.
- If you have nothing to add for this request, return one finding of kind `non-goal`, target `this request`, claim `no agent concerns: <one reason>`, and stop.