Agents, codebases, and teams — Aditya Khandelwal, Amazon AGI Lab
Source
AI Engineer
Author
AI Engineer
Published
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
Why it matters
It is one of the few accounts of rolling coding agents across a whole team rather than one enthusiast, with the org-level changes named. The agent skillA reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.Full definition →-size and context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →-burn heuristics are directly reusable.
Transcript
Leave agent adoption to individuals and the engineer shipping two PRs a day ends up reviewing the ten that the early adopter ships. They fall further behind, the code they are reading is worse, and they conclude the agents are the problem. Aditya Khandelwal's argument, from leading a team of ten through this, is that adoption is therefore a leadership problem and not an IC one, because the changes that actually work, restructuring a codebase for progressive disclosure and converging on a shared setup, are not changes one engineer can make alone. The symptoms of a bad setup are specific. Engineers babysit runs. A simple task burns through 500k of context and hits auto compaction. Someone says the model got dumb today, when the model did not change and the harness did. Their fix centered on one high value skill, called ship it, that carries a change from code done to PR ready, handling the description, the review comments, and CI failures, and often runs for over an hour. That duration scared people until they saw what it bought them. Around it they wired issues and boards into the repo, agentic reviews, and a code gardener that runs nightly. It was not clean: agents filing against each other took the repo to roughly 4,500 open issues in a couple of weeks. Stay for the Q&A, which lands a hard limit near 100 lines in a skill file and offers first prompt context burn as the test of whether progressive disclosure is actually working. Speaker info: - https://www.linkedin.com/in/aditya-khandelwal/ - https://github.com/adityak6798 - https://adityak6798.github.io/ Timestamps: 0:00 - Why solo agent advice breaks on a team 1:30 - The adoption journey, from mandates to slop 2:10 - Fear against confidence, the two axes 4:02 - Is a CLAUDE.md and some skills enough 4:38 - Symptoms that your setup is wrong 5:52 - Why this is leadership's job, not an IC's 6:31 - The review burden trap 7:09 - Harness engineering principles 7:45 - Smart prompt injection 8:25 - Close the loop and keep iterating 9:08 - The playbook: start with the basics 9:47 - Ship it, the one high value skill 11:02 - Winning over the skeptics 12:18 - What went wrong along the way 12:57 - Let prototypes opt out of the standards 13:34 - Stop saying the model is dumb 14:10 - Commit to the turn 14:47 - Q&A: strategies for progressive disclosure 16:03 - Measuring context burn on the first prompt
videoTaking Reinforcement Learning Cross Datacenter — Nan Jiang, Modal
videoEvolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic
videoMultiplayer agentic engineering — Arjun Singh, Superconductor
videoVelocity Sickness: What Happens When Your Whole Team Gets 10x Faster — Matt Dailey, Ref.
