Agentic orchestration
Many agents, many roles, each checking the others.
Models are better when they complement and check one another. Give agents roles, run them side by side on the same branch, and let each one hold the next to account, with people approving at the checkpoints that matter.
Plan: retry budget per tenant + a ceiling · approved
Reaches checkout, ledger · 5 suites to run
Medium · payment path · needs a human approval
retry.go, budget.go · +38 −4 · tests added
2 findings sent back · both fixed by the author agent
216 passed · checkout-smoke passed in Chrome
Sarah Chen approved the change
5 of 14 targets rebuilt · the rest from cache
Landed in order after #417
Shipped to staging · lead time 3h 12m
The roles
A team for every phase of the lifecycle.
01Available
AI design
Investigates the code in plan mode, lays out the alternatives and their tradeoffs, and proposes a plan for you to approve.
02Available
AI impact analysis
Works out which subsystems, services and tests a change reaches, from the graph of your code.
03Coming soon
AI risk assessment
Scores each change for risk before it is reviewed, and says why.
04Available
AI code generation
Makes the change on the branch, with the same tools, permissions and checks as a person.
05Coming soon
AI code review
Reviews every change request and sends findings back to the author agent.
06Available
AI test generation
Writes the tests for its own changes; drafting browser tests from a sentence is next.
07Available
AI test run
Runs only the tests a change affects, and plain-English QA tests in a real browser.
08Available
AI build
Rebuilds only what changed, from a cache the whole team and every agent shares.
09Coming soon
AI merge
Lands approved changes in order once every check and person has signed off.
10Coming soon
AI deploy
Ships to your environments and reports the DORA metrics as it goes.
Built today
Orchestrate the agents you already use.
Define the team
Write the roles your project needs, from reviewer to infrastructure specialist. Agents hand work to each other in parallel or in order, and each runs in the background where you can message it, steer it or stop it.
Any agent, one context
Claude Code, Codex, Gemini and the ReasonOS agent all work in the same branch and query the same map of the code, so a reviewer sees exactly what the author saw.
People at the checkpoints
Agents ask before they act. You approve plans before code changes, roll any AI edit back to a checkpoint, and branch rules keep a person’s approval between a change and main.
AI FinOps
Know what your AI spend buys.
- Every AI call metered, by person, project and model.
- Monthly token budgets per organization, enforced on every call.
- The cost of each change: per task, branch and change request. Coming soon
- Whether the spend pays off: cost against lead time and rework. Coming soon
AI deploy
DORA metrics that report themselves. Coming soon
The same history that holds every change, test run and approval measures how your team ships, with nobody filling in a spreadsheet.
- Deployment frequency
- Lead time for changes
- Change failure rate
- Time to restore service
Evidence of completion
Proof the work was done, and done right.
- Every QA run keeps a screenshot, the logs and a session recording for each step.
- Issues link to their branch, change request and commits.
- The activity log records who approved what, and when.
- One exportable record from the idea to the deploy. Coming soon
Get started
Put a team of agents on your next change.
Open a branch, give the agents their roles, and approve the plan.