mode-classification

Mode Classification

Pick one Agentlas meta-agent mode before generating or repairing files.

Procedure

1. Inspect the user request and any provided path, repo, ZIP, prompt, or agent files.
2. Step 0 - existing material wins: if existing material is being converted, repaired, cleaned, imported, or released, choose agentlas-packager.
3. Step 1 - count independent ownership boundaries. Ask how many roles must independently own all three of:
   • their own memory/context;
   • their own tools/permissions;
   • their own success criteria. One boundary means single-agent-creator. Two or more boundaries means a team-builder candidate. If the boundary count is unclear, run the clarify question loop before generating; do not infer from the word "team" alone.
4. Step 2 - check synthesis need for multi-boundary candidates. If those role outputs must be routed, reviewed, synthesized, or chained through produces/consumes dependencies, choose team-builder and require an orchestrator/HQ plus memory, policy, eval, and QA. If the roles are unrelated, create separate single-agent packages instead of one team.
5. Step 3 - shape guard. single-agent-creator may have many skills/tools but must not emit multiple loose worker agent.md files. team-builder may be small, but it must not omit the orchestrator/HQ.
6. Use keyword signals only as hints after the ownership-boundary check:
   • MULTI hints: separate memory partitions, tools or permissions that must not be merged, role-to-role review/policy separation, and produces/consumes pipelines.
   • SINGLE hints: one coherent job, many tools/skills owned by one worker, no routing or final synthesis requirement.
7. Overlay check: if the request depends on knowledge search over user documents, evidence-based or citation-attached generation, or a document corpus (HWPX/docx/pdf/제안서/계약서/견적서), additionally apply the ontology-backed-agent overlay (modes/ontology-backed-agent.md) with ontology_backed: true on the chosen base mode.
8. Loop policy: derive loop_policy from task purpose and risk using .agentlas/contract-injection-map.json risk tiers — none for simple one-shot tasks, self-correct for complex or long-running work, verified (separate-context verifier + side-effect gate) when the agent performs external writes or sends. Do not force loops onto simple tasks.
9. If the choice changes the output and the request is ambiguous, run the clarify question loop instead of guessing.

Return

Return the selected mode, whether the ontology-backed-agent overlay applies, the derived loop_policy, and one short reason. Then route to the matching builder.

Reference

See docs/mode-classifier.md.