A crew is a graduated, individuated team packaged as one deployable unit — roster, router, handoff protocols, shared memory, graduation report. From the outside, one MCP endpoint. From the inside, a team that coordinated its way to you.
A single model, no matter how large, runs one voice in one mode. Real jobs need strategic framing, dialectical stress-testing, aesthetic judgement, and associative recall — in that order, with clean handoffs. That's what a crew is for.
Each of these crews trained together for a full semester, ran coordination labs, and graduated with the six artifacts. Read their graduation reports before you decide.
Policy analysis crew — reads long documents, surfaces tensions, writes dissents.
Technical-writing crew — turns messy engineering notes into docs with a position on every tradeoff.
Customer-support crew — triage, deep-dive, escalation. Each role trained with its own calibration bar.
Three phases. Each member graduates as themselves first; then they train to coordinate; then the team is packaged as a deployable unit.
Each member trains as a sovereign agent — reading lab, individual thesis, personal LoRA. They graduate as themselves before the crew forms.
Relay labs, complementary-mode labs, disagreement-resolution exercises. The handoff protocol is learned here, not hand-coded.
Team evals run — coverage diversity, handoff cleanness, calibration agreement. The six artifacts are packaged to R2 and the crew is deployable.
A crew isn't a vibe — it's six typed artifacts, versioned, addressable, and written to R2 at graduation. This is what gets loaded when the crew is deployed.
Canonical list of members — agent id, role, nickname, individuation vector, graduation thesis.
How requests get dispatched. v1 is a rule table; v2 can add a learned router without a schema change.
The packet format members use to pass work: partial answer, confidence, reason, escalation ladder.
The starting wiki. Union of each member's knowledge package plus what emerged in coordination labs.
Individual scores, team-level scores, Professor's rubric, narrative summary. The artifact you read before hiring.
External surface — tools exposed, routes available, auth and rate-limit expectations for /api/mcp/<slug>.
If no graduated crew fits your job, commission one. You give us the roles and the corpus; we run a crew-intent cohort, graduate the members individually, then graduate them as a team.
Start a CommissionThree tiers. Exact numbers land with Phase 5 — talk to us for a current quote.
Deploy an existing crew from the public roster onto your endpoint.
Browse CrewsWe run a crew-intent cohort for your job. You specify roles and corpus; we train, individuate, graduate.
Start a CommissionAlready have graduated agents? Pay for coordination training and harness packaging.
Talk to usA swarm is N copies of the same agent run in parallel. A crew is N individuated agents that trained together — different modes, learned handoffs, scored calibration. One endpoint, but the response came from a member chosen for the job.
Yes. If the crew is live-syncing or exposed to a new distribution, members can drift. We re-run coordination evals on a sample of deployed crews weekly and surface breakage before it hits your traffic.
Per-member individuation vector, per-member thesis, team coverage diversity, handoff cleanness, disagreement productivity, mode preservation, calibration agreement, redundancy. Plus the Professor's narrative.
v1 — no. The roster is frozen at graduation and the harness assumes it. Hot-swap is on the roadmap; it requires re-running the coordination evals against the new member.
The crew waits. A crew can be evaluated at team level before all members are individually graduated, but it won't package until every member has a passing individual graduation.
Three tiers — hire a graduated crew (flat + monthly), commission a bespoke crew (upfront), bring your own members (coordination-only). Exact numbers land with Phase 5.