Great vibes and even better tea 🫖 Here’s what I took away: • @romainhuet and the DevEx team shared their thinking on “What Comes After Loops?” One practical detail from the slides: keep a worklog of the goal, current instructions, who owns what, unfinished work and decisions. When instructions change, reconcile them with work already underway. After a context compaction or handoff, read the log and check the actual state before continuing. That gives an agent something concrete to pick up from when the work spans multiple sessions. • @dkundel made a great argument for defaulting to one agent session before building overcomplicated orchestration. Most tasks can be handled perfectly well in one Codex thread. Give it responsibility for the outcome and let it delegate independent work where that helps. The “Coordinate” and “Manager” skills shown on stage made this concrete: split work with clear ownership, let agents resolve dependencies directly, and name someone to assemble and check the result. If two agents are working together, define which parts each can edit and use cloud env so they don’t overwrite each other’s work locally. Also, avoid creating too many org layers and therefore managerial debt • @whoiskatrin talked about what it would take to keep Codex running for 10,000 hours straight. The “keep CI green” example made this tangible. The accompanying AGENTS.md told the agent to run the failing test before and after a fix, keep notes in ci-notes.md, and open a PR explaining what it tested. It also set clear boundaries: don’t disable tests to make CI pass, don’t trigger unnecessarily large CI runs, and leave the PR for review. Give it a way to check its work. Tell it what matters and when to ask for help.