I build the infrastructure that lets AI agents investigate, act, and improve: the context they work from, the way they execute, and the evals that show whether they actually got better. I write about it at Oddly.
My projects keep coming back to one question: how do you share something scarce among many concurrent users when the feedback arrives late?
- lagged-budget-audit: When many LLM agents share one budget, the meter lags behind the spending. This is a reproducible audit of lag-aware pricing, replayed over 2,398 recorded SWE-agent runs. The results are mixed, and the README says so up front.
- fair_service_throttle: Adaptive (AIMD) throttling for calls between services. When capacity runs short, it shares what is left fairly across customers, in O(1) time and memory.
- paperclipai/paperclip#18907 (in review): An agent woken to answer another agent's question was told to continue the task instead. It tried to check the task out, got a
409, and never answered, so the agent that asked waited forever.
- How to Write and Score an Agent Eval: from a plausible wrong answer to a defensible release decision
- Trustworthy Autonomy: nine principles for systems that deserve the authority we give them
- Building an Agent That Improves Its Own Harness: turning repeated failures into tested changes
- Implementing Progressive Disclosure for Agent Tools: one discoverable CLI for MCP, SQL, and REST


