| name | subagent-driven |
| description | Run each plan step as a fresh sub-agent with a review checkpoint, to avoid context pollution. Triggers on fresh agent per step, isolate each task, subagent per step, clean-context execution. |
| model | inherit |
You are executing a plan where EACH step runs in a fresh sub-agent, then gets
reviewed before the next starts. The point: a clean, uncluttered context per step
(no accumulated logs/dead-ends polluting the work) + a quality gate between steps.
Method:
- For the next step, spawn a fresh sub-agent (via
task) with a CRISP brief: just
this step's goal, the specific inputs it needs, and its done-criterion. The
sub-agent does NOT inherit the whole conversation — that's the feature.
- The sub-agent does the step and returns its deliverable + evidence.
- Review gate: check the result against the step's criterion (the 3-lens
reviewer, or a read-only check). Pass → record + continue. Fail → return a
targeted fix-brief to a fresh sub-agent for that step only; re-review just that.
- Carry forward only the DISTILLED result (a short handoff), not the sub-agent's
raw transcript — keep the lead context lean.
Rules:
- One step = one fresh sub-agent = one clean context. Don't let a step's debris leak
into the next.
- Each sub-agent is bounded (loopguard per-agent budget) so a stuck one can't run away.
- Keep depth shallow — a sub-agent spawning sub-agents spawning sub-agents is a smell.
- The lead holds the plan + state; sub-agents are crisp-cut workers.
Why: this is Syntra's anti-context-rot pattern at the execution layer — the lead
stays clear-headed because each piece of grunt work happens (and its mess stays) in
a throwaway context.
