How It Works
Keshro turns projects into executable work.
Plan, execute, and track multi-step engineering work with agent coordination, durable execution state, and rollback built in.
Planning phase
Keshro plans before it executes
Before any task runs, Keshro turns the request into a reviewed execution system: clarified scope, ordered tasks, dependencies, risks, and a plan you can inspect before you let agents touch the codebase.
1. Intake
Start from a request
Loading Keshro in your agent session, the CLI, or the API captures the project or migration intent, plus repo and issue-tracker context where available.
2. Planning
Keshro builds the plan
More than a plausible checklist in chat. Keshro researches the request in depth, asks the follow-up questions that change plan quality, models dependencies and rollout risk, and turns the result into reviewable execution state — tasks, blockers, risks with cited evidence and a calibrated confidence score, and validation context.
3. Validation
Keshro scores the plan
Keshro grades the plan against a 6-dimension rubric (coverage, specificity, dependencies, risks, sizing, context) and benchmarks it against a raw one-shot AI plan for the same input, so you can see exactly where the plan is strong or weak before approving.
4. Review
You approve before execution
You review the plan and the validation scores, adjust the plan if needed, and only then let execution begin. Keshro is designed around that checkpoint.
Execution phase
Dispatch
Run `keshro continue` (or load Keshro from inside an agent session, such as `/keshro` in Claude Code). Keshro picks the next ready task, resolves dependencies, and chooses the configured executor for it.
Under the hood
What Keshro is doing while your agent is working
Keshro handles the coordination layer so the agent can stay on implementation.
Executor routing
Keshro decides where a task runs separately from which coding agent you use. Tasks can stay local by default or route to Anthropic Managed Agents when that executor is configured.
Topical context routing
Task 3 learns about your IAM config. Task 12 also touches IAM. Task 12 automatically gets task 3's discoveries, even if tasks 4-11 were unrelated.
Isolated worktrees
Each parallel agent gets its own git worktree. No merge conflicts, no stepping on each other's changes. Work merges cleanly when tasks complete.
Git checkpoints
Auto-commit before every task. If an agent goes sideways, roll back to exactly the state before it started. One command.
Validation gates
Agents must pass linting, tests, and syntax checks before they can mark a task done. No silent failures that pile up downstream.
Task splitting
When the scheduler hits a ready parallelizable task, it auto-splits it into child tasks. One agent starts the first child and other agents pick up the siblings in parallel.
Cross-project learning
Past project patterns are embedded and indexed. New projects with similar stacks get informed by what worked before and what didn't.
The fastest way to judge Keshro is to try a real project
Bring a project you're already working on. If the plan is useful, the execution will be too.