Sustained runs now up to 14 days
Agents that ship for days,
not minutes.
CDev runs autonomous coding agents on isolated remote infrastructure. Hand one a repository and an objective — it plans, writes, tests, recovers from its own failures, and opens a reviewable pull request when the work is actually done.
SOC 2 Type II · Self-hosted available · Free for open source
- 14:02planMapped 41 call sites for the session-token migration
- 14:09editsrc/auth/session.ts — +182 −64
- 18:44test412 passed, 0 failed, 3 skipped
- 03:17recoverUpstream conflict on main — rebased and re-ran the suite
- 09:52reviewOpened PR #2841 — awaiting human review
- Repository
- acme/payments-core
- Branch
- cdev/oidc-migration
- Orchestrator
- cdev-3.2-long
Running in production at
- Northbeam
- Kestrel Systems
- Ledgerline
- Vantage
- Arclight
- Meridian Labs
- Helix Grid
- Tessera
Built for work that outlasts a single session.
Most coding agents are tuned for one turn of conversation. CDev is built around the run — a long-lived, isolated unit of work with its own state, budget, guardrails and audit trail.
Multi-day execution
A run holds its plan, working memory and branch for up to 14 days. No context resets, no losing the thread between sessions.
Isolated by default
Every run gets a disposable microVM with its own filesystem, network policy and secret scope. Nothing crosses between runs or tenants.
Failure recovery
Flaky suite, broken dependency, upstream rebase — the agent diagnoses it, returns to the last green checkpoint and carries on.
Auditable traces
Every command, diff, tool call and decision is recorded. Replay a run end to end, or diff two attempts against each other.
Parallel fleets
Fan one objective across dozens of agents working separate packages, then land the result behind a single reviewable pull request.
Policy guardrails
Scope agents to paths, commands and dependencies. Require human approval before anything touches production or leaves the sandbox.
Three steps, then walk away.
Starting a run takes about a minute. Everything after that is the agent's problem.
- 01
Connect the repository
Install the GitHub or GitLab app, or point CDev at a self-hosted remote. Read scope, write scope and protected paths are configured once, per repo.
- 02
Describe the outcome
Give the agent an objective in plain language, the command that defines done, and a time budget. It writes and shows you the plan before it writes any code.
- 03
Review the pull request
Come back to a branch, a green suite and a complete trace of how it got there. Approve it, comment inline, or send it back for another pass.
$ cdev run --repo acme/payments-core \
--goal 'migrate session auth to OIDC' \
--verify './ci/verify.sh' \
--budget 72h
run_8f2a1c queued — sandbox ready in 4.2s
streaming trace at cdev.sh/r/8f2a1c
# Detach whenever. The run keeps going without you.
One objective. Sixty-eight hours. Zero check-ins.
An abridged trace from an auth migration across twelve packages — the kind of work that normally stalls for a quarter because nobody wants to babysit it.
- 09:14cloneIndexed 214k lines across 12 packages
- 10:02planSplit the migration into 6 checkpoints, each independently revertable
- 11:36editAdded an OIDC discovery client behind a feature flag
- 17:20testBaseline suite green — checkpoint 1/6 sealed
- 02:41editRewrote 41 call sites onto the new session interface
- 06:08test9 failures in packages/billing — traced to a stale fixture
- 06:22recoverRolled back to checkpoint 3/6, regenerated fixtures, re-ran
- 15:55testSuite green — checkpoint 5/6 sealed
- 04:12editDeleted the legacy session module and its compatibility shims
- 08:30benchp99 auth latency 214ms → 96ms
- 09:47docsRewrote the auth runbook and the migration notes
- 09:52reviewOpened PR #2841 — 34 commits, 6 checkpoints, full trace attached
- Wall clock
- 68h 41m
- Commits
- 34
- Test runs
- 11,208
- Human interventions
- 0
- Autonomous agent-hours
- 4.1M
- Pull requests merged
- 12,400
- Sandbox uptime
- 99.98%
- Longest supported run
- 14d
What teams do with the time back.
“We pointed it at a four-year-old monolith on a Friday and reviewed the pull request on Monday. It had written the migration, the backfill and the rollback path.”
“The trace is the actual feature. I can see why it made every decision, which is the only reason our security team ever signed off on this.”
“It recovered from a broken upstream dependency at three in the morning and just kept going. That is still the part I find slightly unsettling.”
Give an agent something that takes three days.
Pick one repository and one task you have been putting off all quarter. Start the run, close the tab, and read the trace when it lands.
No card and no sales call while you are evaluating