Azeem, as good as these workflows are using the latest agentic softwares, to make this real for people, what might be helpful is to venture an answer to the question: How will this change what my job even is?"(in re Ben Evans’ point that we often ask, "How will this help me do my current job?")
I’ve been exploring with Claude Code the past few months and interested in going to the next level of how to implicate autonomous agents.
It would have been interesting to hear about how much time it has taken you on some of the discussed topics. It seems to me that some of them would have been reasonably light touch to implement, with others requiring deep implication and constant nurturing. For me this is something I’m thinking about as to what is accessible and easily transferable and what requires much deeper implementation time/cost.
The 24/7 angle hits home - I've been running a similar setup where my agent handles tasks overnight without me touching it. What you described with R Mini Arnold maps closely to what I built: specialized sub-agents for different domains, continuous operation, and a lot of course-correcting in the early days. The 179 failures in six days is brutally honest and matches my experience.
The failures are where the real learning happens - each one tells you something about the instruction layer that wasn't explicit enough. I wrote about building the nightshift version of this: https://thoughts.jock.pl/p/building-ai-agent-night-shifts-ep1 - curious whether you ended up hardcoding task boundaries or letting the agent decide scope dynamically.
Azeem, as good as these workflows are using the latest agentic softwares, to make this real for people, what might be helpful is to venture an answer to the question: How will this change what my job even is?"(in re Ben Evans’ point that we often ask, "How will this help me do my current job?")
I’ve been exploring with Claude Code the past few months and interested in going to the next level of how to implicate autonomous agents.
It would have been interesting to hear about how much time it has taken you on some of the discussed topics. It seems to me that some of them would have been reasonably light touch to implement, with others requiring deep implication and constant nurturing. For me this is something I’m thinking about as to what is accessible and easily transferable and what requires much deeper implementation time/cost.
The 24/7 angle hits home - I've been running a similar setup where my agent handles tasks overnight without me touching it. What you described with R Mini Arnold maps closely to what I built: specialized sub-agents for different domains, continuous operation, and a lot of course-correcting in the early days. The 179 failures in six days is brutally honest and matches my experience.
The failures are where the real learning happens - each one tells you something about the instruction layer that wasn't explicit enough. I wrote about building the nightshift version of this: https://thoughts.jock.pl/p/building-ai-agent-night-shifts-ep1 - curious whether you ended up hardcoding task boundaries or letting the agent decide scope dynamically.
Very enlightening.
I thought you stated that you thought the Governance roles in AI currently being pushed will probably be finished in about 5 years.
What AI roles would you recommend for someone beginning or having to change careers?
Thanks, Robert Dudney