FIG. 00 · ARCHITECTURE
A dark factory for software and operations.
Agents do the work end to end. Machines check every step before a person sees it. People decide only where the evidence says they still need to. This is how it is built.
FIG. 02 · THE PIPELINE
Signals in, verified changes out
01 · SIGNAL
Intake
Requests, error reports, CI failures and feedback arrive through connectors, are deduplicated, and become work orders.
02 · STATE
One writer
A work order moves through a state machine with a single compare-and-set writer. Every change records who made it and why.
03 · BUILD
Agent stages
Each stage runs one agent in an isolated checkout with only the tools it needs, and returns a typed, schema-checked result.
04 · VERIFY
Three tiers
Your CI, a browser agent on a preview deployment, and a local harness. A change that fails any of them goes back, not out.
05 · RELEASE
Gated
Merge and release wait for a person until the gate has earned autonomy on measured evidence.
FIG. 03 · TWO PLANES
Your code and data stay on your side
Coordination runs in a control plane we host. Everything that touches your code, data or secrets runs in a data plane, on a runner that connects out only.
| Aspect | Control plane (ours) | Data plane (yours) |
|---|---|---|
| Runs | Work-order engine, scheduler, console, routing | Runner, agent sessions, checkouts, tests |
| Holds | State, decisions, short summaries | Source code, secrets, database access |
| Network | Accepts connections from runners | Outbound only; no inbound ports |
| Contract | Down: the job, its context, its allowed tools | Up: typed results and events. Never source, secrets or rows |
FIG. 04 · SECURITY
Least privilege by default
Tenant isolation
Every row carries its workspace, enforced by row-level security. Two workspaces' code never runs on the same host at the same time.
No standing access
No production database access, no long-lived tokens in an agent's environment, no open egress. Anything broader is an explicit, logged opt-in.
Secrets on the runner
Keys come through a secrets interface on the runner, never into prompts or the control plane.
Scoped model keys
Agents reach models through a gateway with a key minted for one job attempt, under a budget, and disabled when it ends.
Untrusted text is data
Emails, tickets, error messages and repository content are read as data, never followed as instructions.
Memory never crosses
What the system learns about one business is never used for another. Every outside company that sees your data is listed.
FIG. 05 · ROBUSTNESS
Built to stall rather than ship
Agents report, the engine decides
An agent never moves work forward itself. It reports a verdict, and the engine maps verdicts to states.
Level-triggered
A reconciler decides the next action from current state, so a lost event or webhook never strands work.
Durable queue
Jobs run on a queue with leases, retries and idempotency keys, with a concurrency cap per workspace.
Typed contracts
Every result is validated against a schema where it is produced. A step cannot hand off what the next would refuse.
Gates fail closed
When a check cannot run or a result is unclear, the work stops and asks. Autonomy is earned per gate, per workspace.
Every loop counts
A sweep that finds nothing says how many it examined, so silence and success never look the same.
02:14:07 INTAKE 3 error reports grouped into one work order
02:14:09 QUEUE work order claimed · lease 300 s · attempt 1
02:31:52 AGENT fix reported · typed result accepted
02:32:40 VERIFY tier 1, your CI: 418 passed, 0 failed
02:38:11 VERIFY tier 2, preview: checkout flow passed
02:38:12 GATE release waiting on a person
03:00:00 SWEEP stuck work: 0 found of 214 examined
03:12:45 VERIFY tier 1: 2 failed · sent back, not out
FIG. 07 · FLEXIBILITY
Every part behind an interface
Pluggable integrations
Intake, code host, verifier, channels, model provider and secrets each sit behind an interface, even with one adapter.
Where it runs
Hosted in isolated sandboxes, or with the runner in your own cloud account.
Your choice of models
Each workspace allows the model providers it trusts. Routes never fall back to one it has not allowed.
Workers you define
A workspace defines its own workers from tool sets, and an owner approves each grant.
Your own agents
Your assistant can act as you through an MCP server, with your permissions and nothing more.
Your repository
Every change lands as a pull request with its tests, in your repo. Nothing is locked inside the platform.
START HERE
Get your free assessment.
A 30-minute call, then your baseline and what we'd do first, within a week. No cost, no commitment.