P3 · COORDINATE SPECIALISTS

Multi-agent without the cosplay: one speaker, hard budgets, named degradation.

LIVEv1.0.0 · main 88a24f2 · CI greenVerified 2026-08-14

An orchestrator and specialist agents behind a typed handoff protocol, with a global handoff and retry budget, writer-only enforcement on user-facing text, and a crashing specialist that produces an explicit degraded outcome rather than a confident empty one. The evaluation measures routing contracts on deterministic fake specialists — it does not claim several agents beat one model, and the repository says so in its own non-goals. Hire this if you need to be told which part failed.

Open the hosted specialist console Read the source

HOSTED · no clone, no key

curl https://pax-orchestration.vercel.app/health

No clone, no key. Health returned {"status":"ok"} on 2026-08-14. Submit a task in the console, or POST /v1/tasks, and read the ordered trace. Hosted free path on main 88a24f2 / v1.0.0.

LIVE

Every line here is backed by a test, a route, or a file in the repository at v1.0.0 · main 88a24f2 · CI green.

A typed handoff protocol with fake specialists, plus orchestrator-level global handoff and retry budgets.
Writer-only enforcement on final user-facing text, covered by a policy-violation test and ADR 0002.
Degraded outcomes on specialist failure — a crashing Critic and an impersonation attempt both have tests (ADR 0003).
An ordered, JSON-safe multi-agent timeline with completeness, ordering, privacy, and budget-stop tests, downloadable as JSON.
A 12-task golden evaluation across three slices, including an HTTP-absent slice that makes zero network calls.
Hosted console and API on the free path with no key: GET /health and GET /metrics returned 200, and POST /v1/tasks completed with status done on 2026-08-14.
Versioned run replay traces and typed chaos isolation contracts shipped at v1.0.0 / main 88a24f2.

NOT SHIPPED

The boundary, stated before you find it yourself. Nothing below is presented as built, hidden in a footnote, or described as coming soon.

No claim that several agents outperform one. The 12-task evaluation is a routing contract on deterministic fakes.
The hosted specialists are deterministic fakes. The console proves control flow, budgets, and degradation — not answer quality.
No hosted model default, shared memory, or arbitrary tools.
No remote-agent isolation beyond the typed chaos contracts, no hosted quality score, and no cost comparison.
No silent promotion of target architecture into LIVE documentation.

Capture

One image from the repository at the commit named in the caption.

Multi-agent orchestration trace timeline showing handoffs between orchestrator and specialists — The handoff timeline, step by step, with per-agent budget. Source file

INTERVIEW THIS SYSTEM

Three questions worth asking, and the answers I would give.

Published in advance, because a trade-off I can only defend when nobody has read it first is not a trade-off I understand.

01
Why may only the Writer produce user-facing text?
Every extra speaker is another surface that can contradict the final answer or leak an intermediate the user was never meant to read. One speaker makes the output attributable, and the policy is enforced by a test, not a convention.
02
A specialist crashes mid-run. Why is degraded better than retrying until something succeeds?
Silent retry converts a partial system into a confident one. degraded is a terminal state an operator can alert on and a user can distrust appropriately; a retried empty success is indistinguishable from a real answer.
03
Your own eval README says this does not prove multi-agent beats one model. Why ship it that way?
Because the honest claim — that routing and handoffs behave as specified under a budget — is the one the fixtures support. Claiming uplift needs a paired run against a single-model baseline, and that run has not happened.

READ THE PRIMARY SOURCES

Everything above is a summary of a file you can open.

The case study explains why the system is shaped this way; SHIP is the LIVE and not-shipped contract this page is derived from.

Multi-agent without the cosplay: one speaker, hard budgets, named degradation.

LIVE

NOT SHIPPED

Capture

Three questions worth asking, and the answers I would give.

Why may only the Writer produce user-facing text?

A specialist crashes mid-run. Why is degraded better than retrying until something succeeds?

Your own eval README says this does not prove multi-agent beats one model. Why ship it that way?

Everything above is a summary of a file you can open.