Skip to content

Agents

flowchart TD
  C["Coordinator"] --> I["implementer / implementer-haiku"]
  C --> R["package-reviewer"]
  C --> F["final-reviewer"]
  C --> X["researcher"]
  C --> FX["fixers: tier 1, 2, 3"]
  FX --> F1["fixer-sonnet"]
  FX --> F2["fixer-sonnet-high"]
  FX --> F3["fixer-opus"]
Agent Model Effort Tools Job
implementer sonnet medium Read, Edit, Write, Bash, Grep, Glob Builds one plan task, then stays idle for reuse.
implementer-haiku haiku high Read, Edit, Write, Bash, Grep, Glob Builds one Kind: mechanical task as written; stops rather than guesses.
package-reviewer sonnet medium Read, Edit, Write, Bash, Grep, Glob Reviews a checkpointed task or package, then fixes its own findings with regression tests.
final-reviewer opus high Read, Grep, Glob Reads only the review packet; returns a verdict and tiered findings; never fixes.
researcher sonnet medium Read, Write, Grep, Glob, Bash, WebSearch, WebFetch Gathers evidence for one sub-question into the ledger; never writes the report.
fixer-sonnet sonnet medium Read, Edit, Write, Bash, Grep, Glob Tier 1: known cause, one module.
fixer-sonnet-high sonnet high Read, Edit, Write, Bash, Grep, Glob Tier 2: known cause across modules or a seam, or a known probe class.
fixer-opus opus high Read, Edit, Write, Bash, Grep, Glob Tier 3: unknown cause, or a lower tier failed twice.

No agent has the Agent tool, so no agent spawns agents. Where a harness has no plugin agents, each runs as the harness’s generic subagent with agents/<name>.md as its instructions.

Say task T4 of a build fails its tests and a first fixer cannot clear it.

  1. You see an agent named fixer-sonnet start on the failure.

    Claude Code · ~/social
    T4 tests failed
    dispatch fixer-sonnet finding F-T4-1 note .superlite/fixes/F-T4-1.md

    It is tier 1, for a known cause confined to one module.

    The coordinator gives a fixer a finding id, here F-T4-1, and a note path, .superlite/fixes/F-T4-1.md. The note is the fixer’s only memory, because each fixer starts fresh.

  2. The tests still fail. The coordinator moves to fixer-sonnet-high, tier 2, which handles a cause that spans modules or a seam. It fails twice.

    Each failed attempt is written into the note under “Tried and failed”, so the next agent does not repeat it.

  3. The cause is still unknown, so the coordinator sends fixer-opus, tier 3. It reproduces the fault first, then fixes it and adds a regression test.

    It ends with FIXED or NOT_FIXED and a reason. A fixer fixes only the listed finding and never spawns another agent.

  4. Say the coordinator tried to skip a tier on a task marked blocked. The pre-tool hook refuses:

    Claude Code · ~/social
    Blocked: BLOCKED task needs fixer tier opus. Launch the fixer one tier up, not tier T4 need=opus got=sonnet.

    The tier rule is enforced by the hook, not left to the coordinator’s judgement.

What you end up with: T4 fixed by the cheapest tier that could do it, with a regression test and a note that records what each attempt learned. The full loop is in Reviews and fixes.