Agents, codebases, and teams — Aditya Khandelwal, Amazon AGI Lab
Coding-agent adoption is a team systems problem, not an individual tooling choice. Share effective practices, disclose context progressively, and build automated loops that detect and repair weak output.
Khandelwal describes practices developed with a **10-person team**: keep the main agent file as a thin index, connect code to deeper runbooks, add CI and automated review, and continuously feed team findings back into shared skills.
Treat agent enablement as maintained engineering infrastructure. Cap skill entry files around **100 lines**, watch initial context growth, and turn the strongest individual workflows into editable team conventions rather than leaving every engineer to improvise.
Khandelwal describes practices developed with a **10-person team**: keep the main agent file as a thin index, connect code to deeper runbooks, add CI and automated review, and continuously feed team findings back into shared skills. Treat agent enablement as maintained engineering infrastructure. Cap skill entry files around **100 lines**, watch initial context growth, and turn the strongest individual workflows into editable team conventions rather than leaving every engineer to improvise. Automation can create its own noise: one rollout produced roughly **400–500 issues** within weeks, alongside merge conflicts and uneven buy-in. The playbook is one team’s experience, and its thresholds need validation against each repository and workflow.
This moves coding-agent adoption from individual prompting to maintained team infrastructure: thin entry files, linked runbooks, CI, review, and shared skills form a continuously updated operating system for the repository. It reinforces evidence-gated and durable workflow candidates, while the issue flood and uneven adoption narrow the promise: more automation without governance, triage capacity, and environment quality can increase coordination cost.