BacklogJack: AI agent failure replay

Turn a failed agent run into a reusable regression test

Problem

Agent failures are flaky and hard to reproduce: one step produces a bad output but you can't tell which step or why. Developers rerun from scratch instead of replaying a single step and testing a fix. Agentfork and similar show 'replay from any step and test fixes' is a real need — but there's no lightweight version for regular developers.

Solution

Record every step's inputs, outputs, and tool calls into a replayable trace. Pick the failing step, tweak the prompt/args, rerun only that step and downstream, diff old vs new trace, and export the case as a regression test. Works with OpenAI/Anthropic-compatible APIs.

Users

Indie developers and small teams putting AI agents into production workflows

MVP

Thin SDK wrapper to capture traces; web UI to click any step and replay; rerun from a step after edits; old-vs-new trace diff; one-click export as failing test case

Revenue

Open-source core + paid cloud replay and team regression suites; or per-replay metering

Why now

Agent workflows are moving from demos to production, reliability pain is spiking, and debugging tooling is still stuck at 'just rerun it'