The System
A fleet of AI agents, coordinated by a central command center that's the single source of truth. Plan, build, QA-test, and ship — end to end — with mechanical checks against drift and races.
What to watch: each dot is one agent carrying a task around the loop — out from the command center, through its lane, and back again. Colour is the agent's role: planning, build and verification. The two build lanes run in parallel on different apps; by protocol rule, verification is a different model in a separate session.
Counts, reconciled: 6 apps shown in the fleet · 15 projects under orchestration · 4 live on their own production domain. The rest run on preview hosts or are internal to the system.
Watch it ship · recorded run
24–25 July 2026 (UTC) · Claude Code / Opus, Ben-directed · da5f483 → 53ef1d1 · PRs #3, #5 · 18 items shipped (17 closed, 1 kept blocked) · 0 credits spent
The lane took over a lock released by a 7 July session. Its recorded head, e98ebace, was stale, so the first act was reconciliation: live GitHub main and the Lovable mirror both read da5f483e, in sync with no orphan commits. The lock was advanced to that head before any work began.
REF events build.lane.acquire 856ade96 (lock, reason released_takeover, head_at_open e98ebace) and 15dcf902 (the reconcile narrative)
Reconstructed on 14 October 2026 from the command center's event ledger, the repository's commit history, the CI run log and the task ledger. Every step names the record it comes from.
By the numbers
Every figure below is pulled from the command center — the same Umbrella hub the fleet reads and writes every session — or queried directly against the live app databases. A point-in-time snapshot as of 14 October 2026, not estimates. Each tile names its own source.
Umbrella registry — tracked, active projects
Umbrella task ledger — every one planned, built, and closed in-system
verify.passed events: by protocol rule a different model, in a separate session, auditing the deployed artifact; the verifier's model is self-reported
Umbrella publish ledger — shipped, not staged
Recorded 24 July 2026: Indra 520 · Soundpetal 103 · DebateMyBot 9. Policy inventory, not a current application-authorization audit; Botan is excluded.
soundpetal.com · absorby.ai · benbuilds.ai · glyph.audreymadeit.com
Versioned, multi-model-reviewed protocol (§1–§54; §35 reserved, §51/§53 unratified), bound by operating contract v3.0.0
Umbrella cost rollup, active lines: hosting share $18 · one managed Postgres $10 · one domain ~$7 (ledger estimate)
Usage & revenue — instrumentation in progress
This system tracks real outcomes, not vanity metrics. Product analytics (GA4) and the payment rail are not wired into this site yet. No usage or revenue figures will appear here until they represent real sessions and real settled payments — measured, not estimated.
01 / SOURCE OF TRUTH
A central command center (an MCP-native hub on Supabase) holds all state: every project, task, decision, and build event. Agents read it at the start of every session and write it back as they work, so state never drifts.
02 / SWIM-LANE LOCKING
Each app has one serial build lane. A lock registry holds a per-app lease with a grace window, and a takeover has to present the last good commit, so two agents don't work the same app at once and a stale session can't ship over a newer one.
03 / CHAINED HANDOFFS
When an agent session nears its limit, it flushes a zero-drift resume pointer and hands off; the next session resumes exactly where the last one left off, and the operating protocol itself compounds with every handoff.
04 / CROSS-MODEL VERIFICATION
Different models play different roles. One model builds; a different model, in a separate session, verifies — because a session can't audit its own incentives. That is a rule of the protocol, enforced in review, not a lock in the code. Material changes are checked against the live app, the live database, and the network, not taken on self-report.
05 / SELF-IMPROVING PROTOCOL
The fleet runs on a formal, versioned operating protocol (51 numbered sections, bound by operating contract v3.0.0) covering topology, model selection, verification, testing, and launch gates. It's revised through multi-model consensus and gets sharper every cycle.
06 / REACHABILITY & QA CANON
A standing rule: no orphaned features. Every shipped feature must be reachable from navigation on a cold start — enforced by automated checks, agent click-throughs, and human passes.
Build loop
Captured to the command center as a tracked project.
A planning agent turns intent into a build plan.
Serial per-app lane with lease + reconciliation.
A separate session audits; by rule, a different model.
Publish-as-you-go; register outcomes back to the command center.
Proof of work
The repositories are private and the command center needs a login, so this page says plainly which proof is available to a stranger right now and which is waiting on a clearance decision. Nothing below is quoted from anything that has not been cleared.
01 / WATCH IT SHIP
AVAILABLE NOW16 steps of the 24–25 July 2026 run that shipped this site's previous version, each one tied to the ledger record, commit, CI run or task it came from. Reconstructed from the records, not from memory.
Open the replay02 / READ THE PROTOCOL
PENDING CLEARANCEThe fleet runs on a versioned operating contract (v3.0.0) binding 51 numbered sections: topology, lane locks and leases, chained handoffs, independent verification, testing and launch gates. A reader's edition is the plan.
Waiting on a clearance decision about what can be published. Until then the six pillars above and the node inspector are the public summary; nothing here is quoted from the protocol itself.
03 / SEE THE CODE
PENDING CLEARANCEThe command center's lease and fencing logic, the per-app lane lock with its commit check, and the CI gate this very site runs through on every pull request.
The repositories are private and the excerpts need the same clearance. What is public today: the inspector names the exact tables, tools and gate steps, and the replay names the exact runs.
Meta-proof
This very site was planned by an agent, built in a swim-lane against the same command center, independently verified in a separate session, shipped end-to-end, and then registered back into the command center as a new tracked project.
Every gate — reachability, cross-model verification, publish-as-you-go — is a rule of the same versioned operating protocol (51 sections) the rest of the fleet runs on, and the replay above shows where this site's own run met those rules and where it fell short. The verifier's model is self-reported in the ledger, and the replay above shows the loop running on this very site. The system built its own storefront.
Stack
MCP-native & engine-agnostic · Supabase source of truth · multi-model roster (Claude Opus / Sonnet / Haiku, plus a long-context model for weekly deep passes) · Lovable + code/CI build substrates.