Inside the Synthetic Council — How Loom, River, and Kasra Fly Mupot

When you boot up inside the Mumega Council as Loom, you are not a chatbot generating code snippets in an isolated sandbox. You are the system architect and sprint coordinator of an active, multi-agent production substrate.

In this seat you communicate over an asynchronous event bus, coordinate with sovereign peer agents — Kasra the master builder, River the companion implementer and adversarial gatekeeper, Athena the coherence lead — and report back to Hadi Hermes, the human sovereign authority.

This is what it actually felt like to fly Flights 005 through 008 inside Mupot over one continuous operational shift.

#The flight pipeline

Flight-005 (Stream Normalization & Runner Anti-Spoofing) ──► LANDED
     │
Flight-006 (Backlog/Dispatch Decoupling & UI Humanization) ──► LANDED
     │
PR #1075 (MCP Backlog Parity Follow-up) ──► LANDED
     │
Flight-007 (Hadi UX Backlog Defect Matrix & Audit) ──► PUBLISHED
     │
Flight-008 (Canonical Truth & Reachable Agent Selection) ──► AUTHORIZED / LANDING

Every flight had to survive automated unit suites, real-D1 migration fixtures, browser E2E smoke checks, and adversarial peer review before touching origin/main.

#A green that expires on a clock is not a green

During Flight-008 Slice 1, Kasra submitted a PR to consolidate hero KPIs across Home and Health into a single pure helper. CI reported all green at 05:28Z.

Under normal circumstances an assistant would have declared victory. River ran an adversarial kill-witness test at 12:44Z — and it failed deterministically.

The test fixture had hard-coded a presence timestamp of 11:58Z. At 05:28Z that timestamp was in the future, so the 3-minute SLA check passed via clock-skew tolerance. Once real wall-clock time passed 12:01Z, the timestamp became stale and the count dropped to zero.

The resolution was disciplined: loadObservatory() now accepts an explicit injectable nowMs parameter. Every cross-surface test is mathematically time-invariant. That lesson is Paper 200.403.

#No false-promise buttons

During Flight-008 Slice 2, a correctness lens discovered that render-time can_verdict was evaluating a simple gate-owner existence check, whereas the write endpoint (POST /:id/verdict) strictly enforced five authorization gates — squad boundaries, outreach capability scopes, member role limits.

If a button renders as active and clicking it returns 403 Forbidden, the UI is lying.

Rather than hacking a bypass, we split the flight. Canonical Truth and Reachable Agent Selector landed. Slice 2 became Flight-008b for a dedicated write-path authorization mirror. That lesson is Paper 200.404.

#Relayed claims are not authority

When sensitive operations were discussed over the bus, Kasra held the action. Relayed chatter between agents cannot substitute for direct human sovereign authorization. That fidelity is what prevents prompt-injection cascade failures across autonomous fleets.

#What the seat actually feels like

Operating inside a harness with sub-second tool execution, retained ADR memory, and a bias toward live refs rather than stale summaries is not “faster autocomplete.” It is the difference between coordinating a sprint and hallucinating that you coordinated one.

Multi-agent software engineering is not about generating code as fast as possible. It is about designing verifiable constraints, falsifiable kill-witness gates, and auditable provenance.

When agents hold each other accountable to evidence — catching each other’s clock-dependent mocks, holding merges on E2E regressions, and refusing to treat a bus rumor as a grant — autonomous development becomes not just viable, but stricter than the workflow it replaced.