AI Agent Run Debugger

Status: Prototype. New screens for my public transition runtime, Continuity. Data shown is illustrative.

Teams that let an AI agent change code get a result with no reliable account of how it got there: what the repository looked like before, what changed, and whether anything checked it. When a run fails halfway, the evidence is often cleaned up with it.

Continuity is my open-source Python runtime (AGPL-3.0). Each piece of agent work is a state transition in its own Git worktree: observe, plan, execute, observe, verify, then re-observe to check nothing moved. The outcome is journaled before anything is cleaned up. If a step fails, the workspace and the record are kept, and nothing retries on its own. 525 tests pass.

The repo has no interface; these four screens are new. Phase names and journal fields come from the code. kilnworth (demo) and every run, file and hash are illustrative.

Contract & Forward-Deployed Engineering, from $65/hour: an engineer inside your team, from a first block of 10 hours. Forward-deployed and AI systems work, such as getting an agent runtime like this one into production, starts at $90/hour.

Run: one run as a tape of phases, journaled before its worktree is removed.

Ledger: every outcome on one ledger; two of eight task types run today.

Snapshots: two snapshots of one worktree, and the steps that moved it between them.

Retained: a failed run keeps its workspace and its record; nothing retries itself.