Aditya Khandelwal (Amazon AGI Lab) on the part of agent adoption everyone skips: making coding agents work for a TEAM, not just an individual. He led a 10-person team through the full enterprise cycle — solo adopters with outsized leverage → mandates and “token max” → AI slop and sev-2s → retraction — and frames adoption on two axes: fear (am I out of a job?) × confidence (do I actually use it?). The goal is moving everyone to low-fear, high-utilization.
Why the naive setup fails
- CLAUDE.md + skills break in a real team/production context
- It’s a leadership problem, not an IC problem — the most impactful thing an org can do
- Without it: 10-PR-a-day people look like gods, while 1-2-PR people carry the review burden, curse the agents, and never catch up
Harness engineering principles (per codebase)
- Smart prompt injection — treat the whole codebase as a waypoint: docs live in comments, runbooks reflected in code, so the agent finds what it needs when it needs it
- Close the loop — self-healing system; slop is inevitable, build detection + removal into the pipeline
- Iterate continuously — never “done”; the substrate changes constantly
The playbook
- Do the basics: progressive disclosure; find your best ICs’ practices and ship them across as a SHARED setup
- One high-value skill (“ship it”: code-done → PR-ready, handles comments/CI/merge) — often runs over an hour, but it’s the proof the agent can be trusted without babysitting
- Close the loop on process: wire issues/boards into the repo, CI/CD, giant reviews, a nightly “code gardener”
- Win over the skeptics — let them edit and play with the shared setup; that’s the true sign of investment
- Iterate constantly — accept X% of IC time goes to tuning the shared setup
Problems they hit
- 400-500 issues in weeks (agents not wired to issue boards correctly)
- People falling back to babysitting when it’s not perfect — feed feedback into the skill
- Agents taking too long — reframe: good; longer reasoning = better output, and you can go do other things
- Merge hell, slop in experiments (opt prototypes out of rigor standards), people varying on the spectrum day to day
Instead of “the model is so dumb,” ask “how can I make it smarter?” It’s not a personal setup — it’s the shared setup you invest in.
Skiing metaphor to close: commit to the speed or you crash — go full send, be okay with falling, the point is to recover. Q&A adds the progressive-disclosure specifics: skill.md hard limit ~100 lines (the skill is a folder), CLAUDE.md/AGENTS.md as thin index, runbooks in code comments, and the signal it’s working — first prompt grabs ~20-25K tokens, not 40-50K.