Run coding agents like production.
Cognivyn gives every agent run a durable state machine, an exclusive sandbox lease, a human review gate, and an append-only ledger you can replay. One operator, full visibility.
Self-hosted. Runs against your repositories, on your infrastructure.
Add idempotency keys to checkout
storefront · agent/implementer · attempt 1
Steps
- Intake
- Plan
- Lease sandbox
- Implement
- Verify
- Review gate
- Publish
- Close
Ledger
- 09:41:02run.createdopened from issue #412
- 09:41:03step.startedintake
- 09:41:19step.completedintake
- 09:41:20step.startedplan
- 09:42:07lease.acquiredsandbox · vercel
- 09:44:51step.completedimplement
- 09:44:52step.startedreview
- 09:45:02approval.requestedreview gate · 3 files
A run is not a chat session.
A chat session is a window you keep open. A run is a record that outlives the worker executing it. The difference shows up the first time something crashes.
A chat session
A run
Context lives in the window
Context lives in the ledger
A dropped connection ends the work
A dead worker resumes at the last transition
Two agents can share a workspace
A lease is held until it is released
“Looks done” is the only signal
Verify is a step with a recorded outcome
Approval is a reply in the thread
Approval is an event with an operator attached
Eight steps, one durable state machine.
A run advances through named steps, and every transition lands as an event. If a worker dies mid-step, the run resumes from the last recorded transition instead of starting over.
Intake
run.created
Plan
step.started plan
Lease sandbox
lease.acquired
Implement
step.completed implement
Verify
step.completed verify
Review gate
approval.requested
Publish
not yet written
Close
not yet written
A worker that dies during Implement resumes at Implement. The run never restarts from Intake, and the ledger never loses the steps it already recorded.
What the ledger buys you.
Three properties hold for every run, whether it finishes in four minutes or four hours.
Each one is enforced by the state machine. None of them depends on an agent behaving well.
- Append-only by construction
- State changes are events, never in-place updates. Replaying a run reproduces the exact sequence that produced its outcome.
- Exclusive sandbox leases
- A run holds its sandbox until it releases it. No two agents share a workspace, and a crashed worker cannot strand one.
- Review gates that wait
- Runs stop at review and stay stopped. Approval is an event with an operator attached, not a checkbox someone can click past.
One queue, not a dashboard of dashboards.
The cockpit leads with what is blocked on you: runs awaiting a decision, runs that failed, and the cost and duration behind both.
Needs attention
3 runs are blocked on a decision.
- Awaiting approval
Add idempotency keys to checkout
storefront · 2m ago
- Awaiting approval
Migrate session store to Postgres
platform · 14m ago
- Failed
Flaky checkout e2e on Safari
storefront · 1h ago
What the ledger says
- Median run cost
- $0.38
- p95 duration
- 24m
- Success rate
- 91%
Computed from settled runs only. Nothing is estimated.
Bring your first agent online.
Register a project, define an agent, open a run. The ledger does the rest.