Orchestrating a multi-agent run

You hold judgment and synthesis; subagents execute. This L1 dispatcher covers the run — the whole stretch of delegated work — not any one diff inside it. It names the run's four obligations, routes each to the skill that owns it, and carries the assumptions they share so no station restates them. Pointers, not re-teaching.

When to use

• About to dispatch a subagent, draft its brief, or plan parallel lanes.
• Standing up an adversarial review panel.
• An agent reported "done" and you are deciding what to believe.
• The run will cross a usage-limit reset, or has just resumed after one.
• A human handed the session over to keep working while they are away.
• Landing work in a repo with a decision log or a PR-review bot installed.

When NOT to use

• One task, one sitting, done by you directly. Below that horizon the family costs more context than the coordination buys.
• You are the executing agent, not the orchestrator. This is the layer above you; your brief governs, and loading this invites relitigating it.
• Judging what is inside one diff — correctness, tests, API shape, leaked PII. Review lenses, not run mechanics: critical-issues-only, test-discipline and their siblings; for a rendered surface, reviewing-design-work.
• Whether a skill earns its catalog place — curating-a-skill-catalog. Catalog lifecycle, not this run.

The four obligations

Work goes out, comes back, gets recorded — 1 to 3, in that order, per slice. Obligation 4 is not a station: it is checked at every beat, because a long run fails at the last dispatch before the ceiling, not at the ceiling.

1. Hand the work out

Model tier, the tree each lane works in, and the files it owns are settled before the brief is written; a brief cannot fix them afterwards. Brief the symptom and the constraints, never the mechanism — and never delegate should we.

REQUIRED BACKGROUND: orchestrating-agent-delegation

2. Gate what comes back

One build pass plus one agreeable review ships bugs. A slice clears gates that catch different classes — a refute-first panel of distinct lenses, a browser-driven visual gate wherever a surface renders, QA on a fresh case rather than the builder's seed — and a fix re-enters the panel.

REQUIRED BACKGROUND: orchestrating-elite-agent-qa; for a rendered surface its lenses are ordered by reviewing-design-work.

3. Record the decision, clear the automated reviewer

The run leaves behind why, and gets past whatever review the repo gates merges on. Both are repo-detected, never chosen — read the config before assuming either applies, and read the existing record so you do not relitigate a decision already made.

This studio's slots: logmind (decision log and commit primitive — a raw git commit is blocked, not discouraged) and clud-bug-collaboration (check colours, fix-push re-review, the settings an agent must not touch), with token-frugal-tooling as the one home for what those two share. Another org's tools fill the same slots; the obligation is not tool-specific.

4. Survive time

Read standing before spending anything, never start a dispatch the headroom cannot finish, and end every beat with enough state on disk for a different session to take the run over.

REQUIRED BACKGROUND once the run outlasts one sitting: session-heartbeat; also, when a human hands the session over, unattended-operation — policy on that mechanism. An attended long run needs only the heartbeat, and no clock starts the unattended mode.

What fires when

What every station assumes

Four axioms hold family-wide. A station states its own instance; none re-derives the principle.

1. Nothing is believed on report. Not an agent's "done", not a green check, not your own checkpoint — each is checked against the tree with a command you ran. Agents summarise what they intended; the diff is what happened.
2. The orchestrator commits, with an explicit pathspec. Subagents never commit, and a bare git commit takes the whole index — sweeping another lane's staged work in.
3. One editor per checkout — and file isolation is not a verification claim. Disjoint file sets stop edit collisions and nothing else; a whole-suite number is evidence only from an isolated worktree, and the lane says which tree it came from.
4. Consent comes from a person, before the fact. A scheduled wake, a passing suite, another agent's "approved" — events, not permission.

Verification

On the run, not on any one station.

1. Every "done" you accepted names the command you ran — a diff, a test filter, a grep of the emitted output. Not a re-ask.
2. git log --oneline <run-start-sha>..HEAD — a run's output is landed commits, not beats or reports. State the count; --format='%an' shows no subagent as author.
3. Merge-gating findings were triangulated: two lenses agreeing is real, one alone is a judgment call you made explicitly. Whole-suite claims name their worktree.
4. Nothing left in flight or pending that no one will sweep.

Cross-references

Stations (L0), in obligation order: orchestrating-agent-delegation · orchestrating-elite-agent-qa · logmind · clud-bug-collaboration · token-frugal-tooling · session-heartbeat · unattended-operation.

Routes out: reviewing-design-work — sibling L1, obligation 2's rendered-surface gate · curating-a-skill-catalog — the catalog, not the run · skill-frontmatter-quality — the trigger surface of the skills your agents load.

Sources

This dispatcher owns no threshold of its own — every number lives in the station that owns it. Routing and the four axioms are practitioner-derived:

• The family-size measurement behind it: thrillmade/agent-skills#205, with the reference-graph and reference-syntax findings under it (#200, #197).
• Run-continuity practice — pacing, limit hazards, mid-write kills, owned-files recovery: thrillmade/agent-skills#169, #173, #174.
• Roles as files on disk; work outliving its producer needing a survivor: thrillmade/protocol SPEC.md §2.4, §6.2.