Measure a month of coding-agent sessions against shipped work
At a glance
Yus Erdo tracked prompts, tool calls, commits and pull requests to understand how he actually used coding agents. This workflow is listed under developer-productivity, workflow-measurement, coding-agents.
Over 31 days, he logged about 2,300 prompts across Claude Code, Cursor and Codex. He compared exploratory work, structured workflows and the differences between work repositories and personal projects.
How it went
He reports that about five percent of prompts led directly to a commit or pull request; much of the remaining activity was research, planning and debugging.
Worth knowing
This was one developer's observational log. Work repositories involved more explanation and debugging, while personal repositories involved more new implementation.
Save ideas in this browser. Upvotes count across the board.
MCP-connected runtime QA maps critical flows and entry points, then re-tests the blast radius of each Claude Code/Cursor/Codex change against live Postgres/Kafka/APIs/MCP tools—contracts, authz, OWASP, errors—before merge, with self-heal when behavior changes on purpose.
Open-source local procedural memory that learns only after tests pass, then hands Claude Code/Codex/Cursor a short verified skill note on similar tasks (reported token/tool cuts in published benchmarks).
Local gate: one policy set so Claude Code/Codex/Cursor check with Paveo before shell or file write; dangerous actions refused, every decision logged—no network required.
macOS menu-bar app that finds npm/dev servers and daemons started by Claude Code, Cursor, Codex, OpenCode, or T3 Code—including ghosts after the agent exited—and kills the process tree in one click.