What it does
Ouroboros is built around one discipline: an agent should not be allowed to declare victory on its own. Changes pass an interview gate, then staged evaluation, then a budgeted evolution loop that either produces measurable improvement or stops. It runs as an MCP server, so the same loop can back several agent runtimes instead of one.
Key features
- Interview gate that forces intent to be stated before work starts
- Staged evaluation with explicit pass conditions per stage
- Budgeted evolution loop that halts when the budget is spent
- MCP server interface across many agent runtimes
- Audit trail of each cycle for later review
Install
dsh plugin --profile web add ouroborosCompatibility notes
The loop is deliberately conservative and can feel slow on small tasks. Use it for changes where regressions are expensive.
