I keep coming back to this article in the context of agents -essentially we are bottlenecked by agents passing the next actions (the 'monkeys') back to us. This prevents us from actually delegating tasks effectively to agents.
The article outlines the following process:
1. Describe the monkey. The dialogue between a manager and a staff member must not end until appropriate next moves have been identified and clearly specified.
2. Assign the monkey. All monkeys shall be owned and handled at the lowest organizational level possible.
3. Insure the monkey. Every monkey leaving you on the back of one of your people must be covered by one of two insurance policies: (1) recommend, then act, or (2) act, then advise.
4. Check on the monkey. Proper follow-up means healthier monkeys. Every monkey should have a checkup appointment.
When combined with Karpathy's verifiability criteria to assess whether a task actually is a good candidate to hand off (perhaps step 0), this process becomes very suggestive in the context of (Claude code) agents.
Exactly. But not just conversationally built, also conversationally used. UI interactions are seen as just another type of message in a multi-modal conversation.
That's why I encourage people to not just treat the conversation as a way to make the app but actually try using the app declaratively through the conversation.
The article outlines the following process: 1. Describe the monkey. The dialogue between a manager and a staff member must not end until appropriate next moves have been identified and clearly specified.
2. Assign the monkey. All monkeys shall be owned and handled at the lowest organizational level possible.
3. Insure the monkey. Every monkey leaving you on the back of one of your people must be covered by one of two insurance policies: (1) recommend, then act, or (2) act, then advise.
4. Check on the monkey. Proper follow-up means healthier monkeys. Every monkey should have a checkup appointment.
When combined with Karpathy's verifiability criteria to assess whether a task actually is a good candidate to hand off (perhaps step 0), this process becomes very suggestive in the context of (Claude code) agents.