I build aidashos - a local-first operating system for agent work.
It takes a design document you wrote, compiles it into a fixed plan, runs coding agents against that plan in isolated git worktrees, verifies their output with your project's own tests, has a different vendor's model review the diff, and stops at an approval gate before anything merges. Every step writes rows to a Postgres ledger as it happens.
The bet: an agent session cannot resume itself, and this system can resume it. Work state lives in rows, not context windows - so a crashed process, a spent quota window, or a rebooted machine costs one attempt, never the system's state.
- Unattended, repeatedly, accountably. Cockpits (Conductor, Claude Code, Cursor) are for when you are watching. This is for the other half: overnight backlog burn-down, migrations that outlive a sitting, work with an audit trail.
- Subscriptions, not API keys. Frontier agents run through their own headless CLIs on seats you already pay for. A spent window is a scheduling problem, not a billing one.
- Nothing irreversible without a human. Merges, deploys, spend, and external comms fail closed behind approval gates.
- Vendors are config. Implementer and reviewer are a typed pairing in one TOML file, never the same model, with declared fallback seatings for quota outages.
- README - including the honest "what a fresh clone can and cannot do"
- Design trade-offs
- Code structure
- A document driven to a completed work unit
Cash me on twitter: https://x.com/rahulkindarules
Or connect on linkedin: https://www.linkedin.com/in/rahul-nath-753a3052/
Or read my embarrassing blog posts: https://medium.com/@rahul-nath