Harvest

Harvest, operating the product Bitfab, builds replay infrastructure for software teams that ship AI agent features. The product is organized around a three-step loop: capture, emulate, and verify. Capture records production scenarios automatically, where a scenario comprises everything needed to reproduce a single real customer interaction, including a code execution snapshot (inputs, outputs and each step in execution order with timestamps), a database snapshot reflecting the database state at capture time, and recorded responses for external API calls (described as an API sandbox, marked coming soon). The company states recordings are two-way serializable down to datetimes and add no latency to production code.

Emulate lets developers replay a captured scenario locally against the point-in-time database with external calls mocked, or alter a variable such as a different user or order to test a counterfactual. The system detects and instruments side-effecting calls, replays original outputs for safe emulation, and exposes per-span configuration (for example, a decorator marking a call to be mocked on replay). Managed scenario data is described as searchable across large scenario volumes, connectable to monitoring tools, adaptive to code changes, and grouped so similar traces form scenarios that define behavior. Verify compares changes against team-defined outcomes: failing traces can be replayed until they pass and retained as regression tests, full scenario sets can be run as regression suites before deploy, before/after experiments can be compared across scenarios, and token spend and cost per run can be profiled and reduced.

The platform is positioned for use inside coding-agent workflows, with a slash-command assistant interface driven from tools such as Claude Code, Cursor and Codex, allowing engineers to fan out multiple code-change experiments and freeze parts of the call stack when testing a specific change.