Eve
Architect and primary author
A build kit for AI agents that survive a crash mid-task, stop at their budget and touch nothing without your approval.
Facts
- Crash test
- Kill the container mid-task and the run resumes
- Approvals
- Every action set to always, first time only, or never
- Spend
- Capped, and broken down per agent and model
Eve is my product: the layer that lets AI agents do real work for a business without becoming a liability.
The problem
Agents that work in a demo fall over in a business. They lose their place when a container restarts, they spend without a ceiling, and they act without anyone approving the risky parts.
What I built
Durable execution on a Postgres-backed workflow engine: a run that dies mid-task resumes where it stopped, and a durability test proves it on every change.
Every action carries a policy — always, first time only, or never — and every run leaves an audit trail with versioned rules and capped spend.
Running it daily
I run Eve on my own businesses every day, in six roles: product assistant, QA, board reviewer, studio agent, email engine and production guardian.
What this means for your project
If you want agents doing work that matters, the hard part is not the model. It is durability, approvals and cost control — the parts I have already built and run.
Built with
- TypeScript
- Postgres
- Durable workflows
- MCP