You define the job, which is right, but I’d push it one step earlier: the definition should decide whether you even need an agent, some will while others won’t, not just how to fill the loop. Fully deterministic work often doesn’t need one, a script beats an LLM that can drift. The deli and stapler failures both start there, a task too vaguely defined to tell you which architecture you’re building, so people reach for an agent by default. Define it precisely and it tells you how much loop it needs. I build these for clients, and it always starts with the workflow, not the agent.
Absolutely. “The Engineer’s Caveat: Not every one-off question needs a six-part loop. Use the simplest thing that works for quick tasks. But for the recurring workflows you depend on, one out of six isn’t “simple”—it’s a liability.”
One thing I've noticed is that most people try to automate the execution before they standardize the process.
If you can't explain a workflow clearly to a new team member, an AI agent won't magically figure it out either. It'll just make the existing chaos run faster.
The "learning" step also deserves more attention. That's the part that turns a workflow from something you repeatedly fix into something that actually gets better over time. I think that's where the real leverage comes from, not from adding more agents.
Really enjoyed this framework. The "stapler that talks" analogy is going to stick with me 😄
"A stapler that talks" is the best description of most AI setups I've seen this year. Automated the hands, skipped everything that makes hands useful.
The deli counter image is perfect too because it captures something specific about how people adopt AI: they shop for capabilities instead of designing systems. Twelve agents by Tuesday, chaos by Wednesday. I've watched smart operators do exactly this, buy the best models, connect the best tools, and end up spending more time supervising the automation than they spent doing the work manually. The promotion to unpaid micromanager is real and nobody warns you about it on the sales page.
The stapler that talks line is such a good way to describe automation without judgment built in.
Thank you John!
You define the job, which is right, but I’d push it one step earlier: the definition should decide whether you even need an agent, some will while others won’t, not just how to fill the loop. Fully deterministic work often doesn’t need one, a script beats an LLM that can drift. The deli and stapler failures both start there, a task too vaguely defined to tell you which architecture you’re building, so people reach for an agent by default. Define it precisely and it tells you how much loop it needs. I build these for clients, and it always starts with the workflow, not the agent.
Absolutely. “The Engineer’s Caveat: Not every one-off question needs a six-part loop. Use the simplest thing that works for quick tasks. But for the recurring workflows you depend on, one out of six isn’t “simple”—it’s a liability.”
Anyone who’s truly using AI properly is 100% busier as a result! You nailed it
It’s addictive. 😂
One thing I've noticed is that most people try to automate the execution before they standardize the process.
If you can't explain a workflow clearly to a new team member, an AI agent won't magically figure it out either. It'll just make the existing chaos run faster.
The "learning" step also deserves more attention. That's the part that turns a workflow from something you repeatedly fix into something that actually gets better over time. I think that's where the real leverage comes from, not from adding more agents.
Really enjoyed this framework. The "stapler that talks" analogy is going to stick with me 😄
Mission critical. 100%
Realizing you have accidentally become a full-time, unpaid manager for a bunch of scripts is a painful milestone.
Most people stack up tools thinking it gives them leverage, but it usually just leads to a higher software bill and more digital firefighting.
The real shift happens when you stop focusing on the prompt itself and start focusing on the guardrails and handoffs.
Thants it!
"A stapler that talks" is the best description of most AI setups I've seen this year. Automated the hands, skipped everything that makes hands useful.
The deli counter image is perfect too because it captures something specific about how people adopt AI: they shop for capabilities instead of designing systems. Twelve agents by Tuesday, chaos by Wednesday. I've watched smart operators do exactly this, buy the best models, connect the best tools, and end up spending more time supervising the automation than they spent doing the work manually. The promotion to unpaid micromanager is real and nobody warns you about it on the sales page.
I am afraid you have lived it too, just like I did! 😂
People don't need more agents.
They need fewer broken handoffs.
That's where most of the wasted time lives 👊
The agents that keep breaking usually teach the most, so which one has given you the most grief to get working properly?
The real shift comes from creating systems that protect attention, energy, and decision quality.
Interesting paradox, more automation often increases coordination load if delegation boundaries are unclear