Amazon AGI Lab: Adopting AI Coding Agents is a Leadership Problem, Not an IC One
AI Engineer · youtube · 2026-08-11
Aditya Khandelwal from Amazon AGI Lab shares lessons from leading a 10-engineer team through AI coding agent adoption. He argues that leaving adoption to individual contributors (ICs) creates a trap: high adopters ship PRs rapidly, while others fall behind reviewing bloated code and conclude the "model got dumb." Thus, agent adoption is fundamentally a leadership problem requiring collective codebase restructuring.
Key Implementation Practices:
- The "Ship it" Skill: A high-value skill running over an hour that handles transitioning code to a ready PR, including descriptions, review comments, and CI failures.
- Automated Infrastructure: Integrated issue tracking, agentic code reviews, and a nightly "code gardener" around the core skill.
- Pitfalls: Letting agents file issues against each other resulted in roughly 4,500 open issues in just a few weeks.
Engineering Insights:
- Progressive Disclosure: During Q&A, Khandelwal suggests a hard limit near 100 lines for a single skill file.
- Context Burn: Recommends measuring first-prompt context burn as the ultimate test of whether progressive disclosure is actually working.
More from coding & agent
- Anthropic Releases Free Guide: Building Enterprise-Grade AI Agents — goyalshaliniuk · 2026-08-11
- 86 Obsidian Plugins for AI and LLM Integration: A Curated GitHub List — tom_doerr · 2026-08-11
- Llama-CPP Parallel Agents: Prefill Grinds All Other Agents to Halt — EmPips · 2026-08-11
- Ex-PlanetScale Engineer Joins Cursor to Build Cloud Agents — pswider · 2026-08-11
- Demo: Running 5 Parallel AI Agent Environments in Sandboxes — victor_explore · 2026-08-11
- Migrating 5 Apps Simultaneously from Supabase to Neon with Codex — RichardsonDx · 2026-08-11