30 秒でわかる
Theo revisits an argument from Pete, the creator of OpenClaw, that developers should design loops that prompt coding agents. He describes both the increased error rate he saw in early experiments and the review-and-revision workflows that persuaded him to try again.
- 1Design the loopTheo repeats Pete's argument that developers should design loops to prompt coding agents.
- 2Early results were mixedThe Ralph loop helped Theo imagine longer-running work, but he says it also greatly increased the error rate in his changes.
- 3Review and reviseHe began using agents to review code, provide feedback, make changes and trigger another review.
ポイント
- An agent loop can prompt an agent repeatedly rather than relying on a person to write every prompt.
- Theo found that an early loop increased the error rate of his agents' changes.
- Code review, feedback, revision and re-review can form a different kind of agent loop.
- Theo now believes most agent runs may be better driven by such systems, based on his own experience.
チャプターごとに整えた全文を読む
From prompts to loops
Theo describes Pete's proposal and his initial experience with agent loops. He found them interesting but error-prone.
Here's your reminder: instead of prompting coding agents yourself every time, design loops that prompt them. I'd seen loops like Ralph before, and they helped me think about agents doing more over time. But in my own experiments they also greatly increased the error rate of the changes, so I went back to my usual approach.
Changing his mind
Theo experiments with review and revision workflows and uses Hermes to bring him context. He concludes that many agent runs should be driven by systems rather than prompts he writes himself.
Pete was ahead of me. I started getting agents to prompt themselves, setting up systems where they review code, provide feedback, make changes and trigger another review. I also started using Hermes Agent to bring context to me instead of making me go find it.
I've accepted that Pete is right. Based on my own experience shipping code with these loops, I now think the majority of agent runs probably shouldn't use prompts I wrote myself. It's a surprising change in perspective for me, and I have more thoughts to share after trying it.
同じチャンネルの記事
38:47EN
Theo - t3․gg· 2026年10月28日Theo’s Long-Running Agent Experiments: Verification, Compute and CostTheo Browne · Ping LabsA skateboard game and two compiler projects show how repeated checks and flexible compute can sustain agent work—and how expensive those loops can become.
36:07ENTheo - t3․gg· 2026年10月27日Theo on Token Costs, Coding Agents, and the Price of AI CapabilityTheo Browne · Ping LabsA viral estimate of Lauren's AI spending prompts a closer look at cache pricing, benchmarks, and the changing work of coding agents.
30:37ENTheo - t3․gg· 2026年10月26日Theo’s Three Skills for Reviewing and Shipping with Coding AgentsTheo Browne · Ping LabsHow sol-loop, full-send and full-stop encode review, verification and different levels of merge authority.
0:43ENTheo - t3․gg· 2026年10月23日Theo Distinguishes Security Vulnerabilities from Active AttacksTheo Browne · Ping LabsA short explanation of why disclosed software flaws and reported attacks are different kinds of security news.
29:22ENTheo - t3․gg· 2026年10月22日Theo on Security Vulnerabilities, Stolen API Keys and AITheo Browne · Ping LabsA look at the gap between vulnerability disclosure and patching, and a speculative account of why stolen inference keys may be valuable.
31:13ENTheo - t3․gg· 2026年10月20日Theo on Why Smarter Models Help More Than DevelopersTheo Browne · Ping LabsFrom game modifications to compiler experiments, Theo revisits his assumptions about who benefits from more capable AI.