Automate the software development lifecycle
Build features by coordinating agents across your development pipeline, step in when your judgment is needed.
Live feature pipeline
01
Product ManagerProduct investigation
02
Product DesignerUI/UX design
03
Staff EngineerEngineering requirements
04
Software EngineerImplementation
05
Code ReviewerCode review
06
QA EngineerQuality engineering
Checkout recovery is in Product investigation.
A pipeline shaped by your team
Every feature has a route.
This is one example. Define any pipeline you want, with the stages, agents, skills, and rules that fit your team.
- Product investigation
- UI/UX design
- Engineering requirements
- Implementation
- Code review
- Quality engineering
Discover
Product investigation
Understand the problem before touching the code.
Research the product surface, reproduce the behavior, and turn scattered context into a clear opportunity.
Research agent exploringStage 1 of 6: Product investigation. Understand the problem before touching the code.
Human gates
Your judgment has a place.
Every stage starts manual. Make it automatic only when its requirements deserve to decide.
- Manual
- You approve, reject, or steer the work.
- Automatic
- Passing requirements move the card forward.
Implementation
feature/checkout-idempotencyPrevent duplicate checkout charges
Manual stage gate
Waiting for your approval
Agent run finished
Artifacts committed
Waiting for human decision
Changes ready
diff availableDifferent agents. One handoff.
Pick the right tool and model for each stage without losing what the last agent learned.
Use the agents you already trust
Every stage pairs a coding tool with a model and a skill.
Claude CodeCodex CLICursor CLIOpenCodePiContext becomes an artifact
Each stage commits a durable write-up for the next agent to read.
docs/bento/ui-ux-design.mdInteraction statesAcceptance criteriaOpen decisionsKeep one feature across repositories
Frontend and backend worktrees share one branch, one card, and one history.
checkout/webcheckout/apiSecurity model
The sandbox is the boundary.
Agent work stays inside a per-feature environment. Trusted services keep the credentials and publish the result.
one card / one branch / one sandboxIsolated workspaces
Agents receive isolated worktrees, never host SSH keys or host git configuration.
Scoped publishing
The server narrows short-lived GitHub tokens to the repository being published.
Encrypted credentials
Organization credentials are encrypted at rest and resolved only for the owning run.
Run it your way.Self-host with Docker, or share the board while code and agents stay on your runner.