netanel-abergel and tomer-brook (engineering manager, agentic developer platform) tell how monday.com took one coding agent, Atlas, from cloud Claude Code that produced thousands of unwanted PRs to a teammate whose PRs ship. What it took: an eval “council of the wise” (LLM judges plus CI signals), CI failures fed back into the same session and memory, self-testing skills (MirrorD deploys, Playwright, screenshots and video on the PR), and PR guardrails shifted left to spec and pre-commit. Agents are teammates with identities in Slack, monday, and GitHub, an LLM wiki of team context, session retros, and a shared base agent that collects everyone’s lessons. Results: ~15% of PRs come from remote agents and ~50% of those ship untouched. The bottleneck is now review; next are auto-merge and agentic releases. Advice: start with one concrete problem such as bug duty. Brook also dropped sprints for kanban.
Show: ai-engineering-podcast
Concepts: agent-workspaces · ai-sdlc · ai-guardrails · ai-verification · llm-evals · memory-consolidation · llm-wiki · future-of-software-engineering