AI will never account fully for "taste" nor will it ever truly know priorities as those change and evolve faster than foundation models. One thing I've learned as an engineer that "caring" can be toxic and difficult to quantify but no one can argue with the measurement of (daily) use.
Engineers who use their own products are ones who understand the problem space fully and deeply. This is in very rare supply and anything that starts with "Enterprise ___fill-in-blank___" will find that most of the staff do not actually use the very product that they are building.
Care and time and attention go quickly out the door.
Really appreciate this point about dogfooding. You're right. There's a pretty massive gap between building a product and actually living in it day to day.
When you use your own tool, you feel the friction that no amount of user research fully captures. And yes on the enterprise part. It's so easy for teams to lose that direct connection to the problem when they're abstracted away from the end user.
The taste thing ties directly into this. Taste isn't just aesthetic judgment, it's informed by real usage. If you're not using the thing, your instincts about what's good and what's not are just guesses.
Addy, love the point on comprehension debt in agentic coding—I've seen Claude Opus 4.6 handle multi-file refactors but still need human oversight for edge cases. What's your go-to prompt structure to minimize that 20-30% manual fix time? Tried vibe-coding full specs yet? 🚀 #AICoding
Lesson 6 and 4 together describe most retail AI implementation failures I've seen.
Teams add syncs and stakeholder updates when what's actually broken is the ownership boundary - nobody decided who's accountable when the agent surfaces something the business needs to act on.
And slow decisions follow because the interface was never defined, not because people weren't communicating. More meetings, same buried constraint.
The teams and trust lessons hit different when you're orchestrating AI agents instead of humans. Similar principles though - clear boundaries, explicit expectations, progressive autonomy.
I've been running an AI agent system for months and the systems thinking from big tech applies directly. Especially around failure modes and rollback strategies.
Another well-timed piece, for sure.
AI will never account fully for "taste" nor will it ever truly know priorities as those change and evolve faster than foundation models. One thing I've learned as an engineer that "caring" can be toxic and difficult to quantify but no one can argue with the measurement of (daily) use.
Engineers who use their own products are ones who understand the problem space fully and deeply. This is in very rare supply and anything that starts with "Enterprise ___fill-in-blank___" will find that most of the staff do not actually use the very product that they are building.
Care and time and attention go quickly out the door.
Really appreciate this point about dogfooding. You're right. There's a pretty massive gap between building a product and actually living in it day to day.
When you use your own tool, you feel the friction that no amount of user research fully captures. And yes on the enterprise part. It's so easy for teams to lose that direct connection to the problem when they're abstracted away from the end user.
The taste thing ties directly into this. Taste isn't just aesthetic judgment, it's informed by real usage. If you're not using the thing, your instincts about what's good and what's not are just guesses.
I really enjoyed this article! Which is a good reminder of what I want to look for in my next job.
Addy, love the point on comprehension debt in agentic coding—I've seen Claude Opus 4.6 handle multi-file refactors but still need human oversight for edge cases. What's your go-to prompt structure to minimize that 20-30% manual fix time? Tried vibe-coding full specs yet? 🚀 #AICoding
Lesson 6 and 4 together describe most retail AI implementation failures I've seen.
Teams add syncs and stakeholder updates when what's actually broken is the ownership boundary - nobody decided who's accountable when the agent surfaces something the business needs to act on.
And slow decisions follow because the interface was never defined, not because people weren't communicating. More meetings, same buried constraint.
The teams and trust lessons hit different when you're orchestrating AI agents instead of humans. Similar principles though - clear boundaries, explicit expectations, progressive autonomy.
I've been running an AI agent system for months and the systems thinking from big tech applies directly. Especially around failure modes and rollback strategies.
How autonomy works in practice: https://thoughts.jock.pl/p/my-ai-agent-works-night-shifts-builds
Did Google's internal tools influence how you approach AI agent orchestration?