Claude Code team's Addy on how to maintain critical thinking when using Agents
indigo relays the views of Claude Code team's Addy Osmani, pointing out that the stronger the model, the easier it is for people to let their guard down in a comfortable state where Agent outputs are mostly good and multiple Agents run at once while there is only one pair of eyes. The article divides this cognitive decline into three steps—offloading, deferring, and surrendering—and proposes placing judgment at both ends of the task: beforehand, clearly write out assumptions, goals, boundaries, and completion criteria; during, have the Agent output intermediate artifacts step by step and verify them; afterward, look at the evidence first and then the summary, and use another model to cross-check execution documents and logs. The final threshold before delivery is whether you can explain the output and are willing to sign your name and take responsibility for it.
Why read this
AI score 50
The original article breaks the cognitive decline when using Agents into three steps—offloading, deferring, and surrendering—and offers a response framework of defining beforehand, verifying during, and reviewing afterward.
Topics
View original posthttps://t.co/2gyp2Erpxx
View original
https://t.co/2gyp2Erpxx
Body · AI translation
When an Agent does the work for you, what should we be thinking about? This piece, shared by Addy from the Claude Code team, uses coding as an example, but the method can be extended to any Agent work. The stronger the model, the harder critical thinking becomes—and in ways different from what people expect. Obvious errors become rarer, most outputs are good, and you may be running multiple Agents at once but have only one pair of eyes; in this comfortable state you let your guard down, so critical thinking has to be deliberately scheduled 👀 First, beware of cognitive slippage:
- Offloading: you hand the task to AI, but the result is still your responsibility; - Delay: tool output becomes your next input directly, no longer a hypothesis; - Surrender: AI output quietly becomes your own view, and you don't feel the need to check it;
People slide down step by step, each step small and seemingly reasonable, so few notice. If you can't explain the answer even with the help of an agent, you have already given up ownership. Therefore, put judgment at both ends of the task:
1. Define beforehand: first write down your own assumptions and clarify the goal, boundaries, and completion criteria. For any architectural and execution work, I will complete the design specification together with the Agent. That way both of us know what will be done next and what the completion criteria are.
2. Execute in progress: the Agent generates and iterates, and verification enters the execution loop. Have the Agent output intermediate artifacts step by step, so we can confirm whether parts of the execution results are correct. If it is continuously effective, you can proceed with confidence;
3. Verify afterward: look at the evidence first, then the summary, and review hidden assumptions and decisions. Be sure to verify the Agent's execution results; you can use another model to read the execution documents and the intermediate log records for two-way verification.
The final threshold before delivery: can I explain it, and am I willing to sign my name and take responsibility for it? Output should not exceed review capacity; keep practicing the skills you don't want to lose!