RadarRADAR

We're getting pretty good at making coding agents do things. The harder question is whether the things we're doing actually make them better. Next.js is now publishing agent evals that measure success, cost and time, even comparing results with and without AGENTS.md. This week, Why AI Needs Evals, Not Vibes looks at why measuring agents matters, while Build an Eval for Your Coding Agent puts it into practice.

We're also questioning what we give agents in the first place. The Agent Doesn't Need to Know Everything looks at giving agents the right information when they need it, and Measure Whether AGENTS.md Actually Helps tests whether those extra instructions really improve results. You can go further with planning and context in our Product Engineer role.

Finally, as coding tools give us more choice over which models do the work, One Model vs Model Routing tests whether choosing the right model for each task can beat sending everything to one capable model.