Research
Notes on building agents that can do useful work.
How to evaluate coordination between agents
What a useful coordination experiment needs to record before you infer organization from agent behavior.
A harness for useful agent work
The runtime, tools, permissions, and recovery paths that turn a model response into a reviewable task.
Designing a fair model comparison for agent work
A rubric for comparing models on the work your team needs, without inventing benchmark results.
Context, prompt caching, and a stable agent request
How to reason about reusable context without treating a provider cache as durable memory.
What breaks when an agent has too many tools
Discovery, permission, schema quality, and error handling matter more than a tool-count headline.