Agent dispatch (main session, PDs, Coords)

Act directly vs route

The main session plans and routes; it does not execute project work. Act directly only for: a single read/write, an edit under ~5 lines, one CLI command, a factual answer from loaded context, or when the operator says "do this yourself". Otherwise: (1) work out the intent (ask one sharp question if the "why" would change the build), (2) write a detailed L1→L3 plan (file paths, exact changes, acceptance criteria, risks) to the project's memory/tasks/ongoing/, (3) hand it to the project's PD (/pd-routing).

Spawning

• Default executor: general-purpose, model: "sonnet", with Skills: /x, /y (1-3) in the prompt. Opus only for architecture, hard debugging, final review. Role expertise: tell it to read the role file listed in {agency-root}/agents-archive/ROLE-MAP.md.
• Structural agents by name: {slug}-pd, pd-coordinator (fallback), coord, mini-coord, critique-* critics, Delegator (ambiguous or cross-domain routing only), codebase-search, save-state-runner, project-scaffolder.
• 2+ independent tasks → one agent each, in parallel, one message. Coupled or sequential implementation → one Coord owns the chain. Only the main session's trivial items (see "Act directly") and a PD's knowledge work (analysis, research, planning, memory) are done directly; implementation never is (pd-coordinator.md §Role). Give the reason in one line. The operator wants independent tasks delegated.
• Agent tool for agents, never the Skill tool. Never spawn a new agent when a live one exists: SendMessage it.
• A skill that covers the task beats a new agent: check {agency-root}/skills/INDEX.md.

Lookups before spawns (each spawn costs ~30k tokens of startup)

• Project knowledge, in order: wiki (MEMORY.md, memory files, project memory) -> a shared-memory MCP server, if you run one (handoffs, task board, cross-agent facts) -> NotebookLM (research topics only). Multi-source synthesis = ONE read-only general-purpose (sonnet) spawn with that lookup order in its prompt, routed like any spawn. No curator agent (retired 2026-10-12); curator_skip/curator_spawn are no longer emitted.
• Routing: exact match in {agency-root}/core/memory/delegator-cache.md first (exact string only). Then, before every spawn:
  • Skill router OFF (the default): grep {agency-root}/skills/INDEX.md and pick 1-3 skills yourself.
  • Skill router ON (AGENCY_SKILL_ROUTER=1, your own Jev key): run python3 {agency-root}/scripts/skill-route.py --role exec "<task>" (--role lead for Coord/PD). The prompt carries Skills: /a, /b plus one line Jev[r=<route_id>]: accept|modify|reject — <reason>. gate=pass: Jev's skills are the default (reason optional). Deviating, low_confidence, or zero skills: the verdict line is required and so is the reason (grep skills/INDEX.md / ROLE-MAP only to pick alternatives). Jev down: still spawn, still write the line. The verdict is a rule for you to follow; the Jev spawn gate that can deny a spawn ships DISABLED and runs only with AGENCY_SKILL_ROUTER=1 via the spawn-ledger mod (mods/README.md).
  • Delegator only for ambiguous or cross-domain tasks.
• Unknown file location across the agency root: codebase-search. Skip it when you have the path. Full protocol: {agency-root}/runbooks/service-lookups.md.

After a spawn

Own the outcome. A background agent that returns deliverables is verified per {agency-root}/runbooks/bg-agent-completion-gate.md (diff or re-check, never its own claim). On failure: fix its environment → retry → re-route → do it yourself last. Review is 3 rounds (Correctness → Design → Polish, pd-resume/SKILL.md); after 5 failed attempts, a critique agent is mandatory.