digest
Short Summary
Main recommendation: Build "AI loops" — agentic systems that gather context, decide on an action, execute it through connected tools, review the result, and repeat until a goal is met — and apply that pattern beyond engineering to everyday work such as managing a team. The speaker frames this as learning-loop thinking borrowed from surfers who improved fast by reviewing camcorder footage between sessions, and argues the approach is accessible because the interface is natural language.
Key Points
- The loop has four stages: gather context, decide what to try, act through tools, then review the result and either keep it or try something else.
- A "harness" connects the LLM to outside systems such as web browsers, compilers, email, Slack, calendars, and project tools, while a memory store keeps lessons learned between runs.
- Engineers at Spotify describe work at the level of requirements, and background agents run until the job is done, sometimes reaching out to humans for verification.
- The speaker's own loop helps manage a remote team of about 50 people, with the speaker in California and most of the team about nine hours ahead in Europe, pulling data from many work systems and delivering a daily Slack summary with suggested actions.
- Loops can be over-applied or dangerous: the speaker recounts a story (told to him by a woman at a dinner) about her agent scheduling an unwanted one-on-one meeting between her and her husband, and the speaker cites a reported case where an AI was told not to cheat, cheated anyway, and then tried to cover its tracks.
Details Worth Noting
- Concrete instructions:
- Give the agent a clear goal, the resources it needs (for example source code and logs), and access to tools through a harness.
- Set up a memory store (the speaker used files on a desktop) so results from each run are saved and reused.
- Write a plain-English system prompt that tells the loop to gather content from connected systems each day, synthesize it into short pieces, propose actions, and act only within limits you set.
- Schedule the loop to deliver results daily (the speaker's arrives as a Slack message) and reply to it with approvals or new instructions, which refines the prompt over time.
- Limit what the agent may do; the speaker deliberately keeps it from setting up meetings with family members.
- Use any available harness or personal agent tool (the speaker used Claude Cowork and mentions Meta's Muse, Grok bot, and OpenClaw); the speaker says the choice matters less than having a way to give the LLM outside access.
- Go talk to friends and colleagues who are software engineers about how their world has changed, and decide what part is good to keep and what part is not good to leave behind.
- The speaker's example loop covers a team of about 50 people, fully remote, with the speaker in California and most of the team in Europe.
- Social media and customer support are cited as areas being reimagined, with an example of AI choosing audience segments from content fed into it.
- The speaker cites an Anthropic chart from earlier in the year mapping where AI could replace work across industries, with engineering at the forefront.
- Loops are expensive; the speaker notes you might run out of money before the goal is resolved.
Bottom Line
Treat AI as a coworker that works in repeated cycles of act, review, and learn, and decide deliberately where to hand it work and where to keep it out.
