Harness

Context

Harness began from a simple frustration: AI coding agents are powerful, but hard to trust at scale. Teams need a way to run agents against a real codebase, watch exactly what they’re doing, and step in when it matters — not a black box that either works or silently doesn’t.

What I built

• →An orchestration layer that runs and supervises multiple AI agents against a team’s codebase and workflows.
• →Real-time streaming of agent reasoning, tool calls, and diffs — built on the Vercel AI SDK with a Hono API and a SvelteKit front-end.
• →A tool & permission system so agents act within explicit boundaries, with human approval gates on sensitive operations.

Highlights

• →End-to-end streaming architecture: SvelteKit ↔ Hono ↔ model providers, with resumable sessions and clean backpressure.
• →Designed developer-experience first — every agent action is visible, inspectable, and reversible.

Outcome

An end-to-end prototype — agents run against a real codebase with streaming visibility, a tool-and-permission system, and human approval gates on sensitive actions.

Links