Summary
At the AI Engineer conference, Aditya Khandelwal from Amazon AGI Lab delivers one of the more practically grounded talks on a problem most engineering leaders are quietly wrestling with: how to get an entire team — not just individual standout engineers — to actually ship together with AI coding agents. Drawing on months of hands-on experience leading a 10-person team, Khandelwal maps the characteristic arc organizations follow: from initial fear and low utilization, through company-wide token-maxing mandates and the inevitable wave of AI-generated slop, to a more measured and durable adoption pattern.
The core of the talk is a three-part technical framework for setting up codebases to work with agents at team scale. First, smart prompt injection: structuring the codebase so agents can autonomously find the context they need — documentation living in comments adjacent to the code that references it, for example — without requiring constant human guidance. Second, self-healing pipelines: accepting that some slop will always seep in and building automated systems to detect and remove it continuously. Third, treating the whole thing as ongoing infrastructure work rather than a one-time setup project.
Equally important, Khandelwal argues, is treating AI adoption as a human and leadership problem rather than a tooling problem. Fear is real, mandates backfire, and structural changes to codebases require leadership sponsorship to actually happen. He closes with a practical playbook: identify top individual contributors and codify how they’re making agents work, use progressive disclosure to spread those practices across the team, and establish clear feedback loops so the setup improves continuously as models and tools evolve underneath.
📺 Source: AI Engineer · Published August 11, 2026
🏷️ Format: Deep Dive
