A practical eval loop for coding agents starts with codebase traces and user feedback
EdenEmarco177 · x · 2026-07-23
A dev workflow for building evals around coding agents is being described as a loop: feed the agent the codebase and real traces, iterate on the eval direction with the user, build evals in Harbor, run them, review results, and repeat.
Flow
- Give the coding agent the codebase plus real traces
- Iterate with the user on what to evaluate
- Build evals with Harbor
- Run the evals and inspect results
- Repeat the loop with user feedback
More from coding & agent
- A new app generates 2–10 prototypes from different AI models at once — alexmacgregor__ · 2026-07-23
- Claude Code CLI adds crash-resume support for long-running agent workflows — arthurcolle · 2026-07-23
- Raft 1.0 launches a shared workspace for AI agents, built by Moonshot Kimi CLI creator RC — CodeByPoonam · 2026-07-23
- A Rust word-cloud generator was sped up from 100 ms to 16 ms for no real reason — minimaxir · 2026-07-23
- One in seven MCP registry source repos is no longer publicly reachable — mcpindex · 2026-07-23
- TerraLingua opens public access to autonomous agents in a persistent world — kenneth0stanley · 2026-07-23