The model decides. The runtime executes. The event store records.
Control Room runs teams of AI agents under explicit governance, so every action is recorded and every "done" has evidence.
State lives in Postgres and history in events. Every run survives a restart and is inspectable.
Status board
snapshot · 13 Oct 2026A dated snapshot, counted from the AOS repository by command. Not a live feed.
120 of 131tasks verified
120 of 131 tasks verified, 11 not yet verified
- 3 implemented
- 5 in progress
- 2 backlog
- 1 deferred
- requirements across 22 domains
- 199
- architecture decision records
- 26
- test functions in 149 files
- 3,979
Workforces
- Website workforce24 roles from intent to owner acceptance.Running
- Marketing workforceCore roles verified. Live channel checks in progress.VerifiedIn progress
- Software delivery workforcePaused at an accepted rehearsal.Paused
Three familiar failures, three recorded answers.
Agents act without an audit trail.
Typical run: trace: empty
Under Control Room: Every call has its intent saved before it and a receipt after it.
Retries duplicate side effects.
Typical run: retry 2 of 3
Under Control Room: Unknown outcomes are held, never auto-retried.
"Done" arrives without evidence.
Typical run: status: done
Under Control Room: There is no done state. Work is verified only with passing tests and recorded evidence.
Pain-side details are illustrative states.
One unit of work, start to finish.
Anything that cannot be undone waits for the owner.
Every grant is scoped, counted and expiring.
scoped · counted · expiresEvery spend is checked before it happens.
runtimeProducer and reviewer are different identities.
producerreviewer
One unit of work: intent saved before the external call, the call executed by the runtime, a receipt recorded before applying, review by a separate identity, an owner gate for anything irreversible, then applied. Done is struck; the work is VERIFIED with tests and evidence.
Illustrative lifecycle. The states are real; this run is not a recording.
Every request runs the same loop.
This is how Drevan’s own work runs, session after session, and it is the loop a client project goes through. Counts as of 3 Nov 2026.
- 01A brief before any workEvery substantial request becomes a written brief: the real outcome, context pulled from the project’s own files, a research plan, which specialist agents to use, the quality bar and the risks. The owner picks the depth before anything runs.
- 02Context that carries overEach repository keeps a map of its own code, and a library of 2,300+ skills and agent roles is searched before work starts, so a new session begins where the last one stopped instead of from zero.
- 03Ledgers, not memoryEvery assumption, gap and decision is written down as it happens. Control Room’s own ledgers hold 113 gaps and 250 assumptions; an assumption that proved wrong stays on the record with what it cost.
- 04Evidence before “done”There is no DONE state, only VERIFIED with passing tests and an evidence path. A separate critic reviews the work, and 29 architecture decisions are on record so nobody changes the design silently.
Example brief · scripted, not a live session
Example: a one-line request turned into a brief
Request
“Rebuild our clinic’s appointment booking so patients stop calling to reschedule.”
- Outcome
- Fewer reschedule calls; no double-booked slots.
- Context
- Current booking flow, call logs, the clinic’s existing calendar.
- Research
- How three comparable clinics handle self-service rescheduling.
- Agents
- UX Architect · Copywriter · Frontend Engineer · Critic
- Quality bar
- A patient reschedules in under a minute on a cheap phone.
- Risks
- Patient data stays in the clinic’s accounts; SMS consent.
Only then do the agents start. Each result is checked against the quality bar above, and what was learned goes back into the ledgers for the next session.
Three workforces, three states.
Website workforce
October 2026Running- intent
- gather
- direct
- owner gateowner
- produce
- build
- prove
- owner gateowner
- document
- 24 roles.
- Builders write only inside a disposable git worktree. Writes outside were verified blocked, and there is no shell.
- AOS makes every commit.
- 50 numbered quality checks run before a critic reviews screenshots: anti-slop lints, launch-hygiene checks and an external design detector.
Marketing workforce
- strategy
- targeting
- brief
- concept
- copy
- media
- measurement
- Core verified: seven specialist roles, separate reviewers, bounded grants.
- In progress: live channel connection checks and validation on a real campaign.
131 automated tests pass · counted 13 October 2026
Software delivery workforce
Durable, bounded coordination of coding agents, with independent verification and a local preview.
Paused at an accepted rehearsal.
Verified today, module by module.
task titles from the repository · 13 Oct 2026
- Foundation
- Typed IDs, an error taxonomy, an append-only event store with replay, Postgres with optimistic concurrency, a durable workflow engine with checkpointing.
- Organization
- Organization, department, team and position entities; four separate graphs for reporting, communication, escalation and authority; the strategy hierarchy and the why-chain.
- Governance and security
- The constitution as enforced, versioned policy; green, yellow and red approval gates; a permission boundary; prompt-injection defence; sandboxed code and browser work.
- Economics
- A wallet hierarchy with downward-only allocation, spend limits checked before spending, reservation-based prepaid spending, a full transaction ledger and ROI attribution.
- Workforce lifecycle
- Agent identity and lifecycle; governed hiring; onboarding, probation and certification; handover, succession and retirement.
- Memory and knowledge
- Seven-level memory (L0 to L6), budgeted context assembly, typed evidence-backed knowledge, and lessons extracted from failures and propagated.
- Observability and evaluation
- A black-box recorder with cost forensics, step records and a run timeline, an invigilator with six scopes, evaluation runs and metrics.
- Evolution
- Change proposals as governed artifacts, isolated experiments, canary, promotion and rollback, with the execution loop kept separate from the evolution loop.
- World awareness
- A source registry and scan scheduler, relevance filtering, re-verification of stale knowledge, tool contracts authored from discovered services, bounded repository scouting.
What comes next.
Taken from the repository backlog and the work in progress. Not a projection.
- Website workforceNextA web-reading scout and asset gatherer, the experience layer (vibe recipes, media handoff, licence ledger) and Docker-sandboxed builds.
- Marketing workforceNextQualify live ad-channel connections, then validate on a real campaign.
- Software delivery workforceNextFrom an accepted rehearsal to a first real delivery.
- External-agent boundaryNextMore workers behind it, starting with a second coding-agent vendor. Deferred today.