You're right that harness engineering is separate from prompt engineering - one defines the environment, one guides the task. Most tutorials conflate them, which explains why so many first agents feel unreliable.
I can't explain why, but the Stripe minions stuff is really exciting. Maybe because they haven't been super public with their AI use yet. I love to see big orgs able to adopt their tooling
Really nice synthesis of the emerging best practices, thank you for this writeup :) I feel like "the shape of the thing" is capturing this moment in time well - the slow emerging clarity on how to put the pieces of the puzzle together.
Very well done summary and a refreshing middle ground read amongst those either deep into technical jargon and those on the other spectrum of pure hype. Great work!
Something I noticed after 1000+ agent sessions: the scaffolding you describe is exactly what CLAUDE.md does in practice.
Pre-loaded codebase conventions, acceptable risk ranges, escalation paths - my file grew to 400+ lines and it's what lets my agent make commits without asking on every decision (architecture breakdown: https://thoughts.jock.pl/p/how-i-structure-claude-md-after-1000-sessions).
You're right that harness engineering is separate from prompt engineering - one defines the environment, one guides the task. Most tutorials conflate them, which explains why so many first agents feel unreliable.
Fascinating concept, combining systematic thinking with creative problem solving is the future.
I can't explain why, but the Stripe minions stuff is really exciting. Maybe because they haven't been super public with their AI use yet. I love to see big orgs able to adopt their tooling
yeah, I think if a company like Stripe can do it I’m pretty excited to see how agent native workflows evolve at big companies
Same here!
Really nice synthesis of the emerging best practices, thank you for this writeup :) I feel like "the shape of the thing" is capturing this moment in time well - the slow emerging clarity on how to put the pieces of the puzzle together.
thanks Sandra! We should catch up the next time you’re in SF
Very well done summary and a refreshing middle ground read amongst those either deep into technical jargon and those on the other spectrum of pure hype. Great work!
Definitely appreciate the lack of "pure hype" here
thank you! That middle ground is what I always try to aim for so appreciate the kind words