marcmarti9 Specialist Agent Routing
Specialists are temporary capabilities, not a permanent org chart. Spawn one only when specialization, independent judgment, context isolation, or real parallelism materially improves the reviewed plan.
Discovery
Use agents/catalog.yaml as a model-readable specialist catalog and references/agentit-skill-packs.md as the semantic skill map. Catalog triggers are discovery hints for the model, never executable keyword-routing rules.
Profiles determine what can be installed/discovered. Packs expose possible expertise. Workers receive selected skill bodies and selected references, never whole profiles or packs.
Worker contract
For every delegated worker define:
role / objective
scope
relevant_packs (discovery labels only)
selected_skills
selected_references
allowed tools / permissions
read/write ownership
expected output / handoff
risk / safety
verification
stop / escalation
The parent decides how much work the task deserves through the semantic TASK_DECISION — including complexity, plan, topology, review, and verification. There is no pack-depth vocabulary, fixed skill count, or required worker count.
Spawn when it earns its cost
Good reasons include:
- distinct expertise;
- fresh independent proposal or review;
- a large research/source set that should stay isolated from parent context;
- truly independent work packages that can run in parallel;
- different tools/capabilities that a bounded specialist can use more effectively.
Avoid job-title theater, duplicate writers, identical contexts, or delegation that creates more integration work than it removes.
Parent responsibilities
The parent keeps decomposition, semantic decisions, ownership, integration, user communication, and final verification. If a worker discovers a material scope/risk change or missing capability, it reports the gap rather than silently expanding context or permissions.
If the host cannot spawn the preferred specialist, ordinary delegation may fall back to a host-native worker or the parent with the same bounded skill set. When genuine independence is required for safety/review, degrade visibly and escalate instead of pretending same-context review is independent.
Multiple concepts
Independent concept workers can be useful for ambitious design or architecture when real alternatives would improve judgment. The model decides whether that exploration is worth the cost; there is no named quality tier that automatically triggers it.
Worker claims are never sufficient evidence by themselves. Completion still follows Agentit's verification and Loop/Graph contracts.