I run LLM agents in production every day, at the scale of a small organization.
Most of what I build sits between the model and what ships: deterministic quality gates, regression detection, cost ceilings, and recovery paths for long-running agent fleets. The failures that cost me the most were never in the model. They were in the harness around it.
I write those failures up two ways. As minimal runnable reproductions, and as preregistered measurements.
One file. No build tool. Run it, watch it break, then watch the same run with the fix.
- thread-pool-exhaustion-lab A slow dependency takes down endpoints that never touched it. Then the same run behind a bulkhead, so you can see what isolation buys and what it costs.
- duplicate-coupon-lab A per-user limit sits right there in the code and the same person still gets two. Then the same run with the check and the registration fused into one operation.
- rdb-join-lab Benchmark harness for a preregistered study of ORM query patterns: JOIN vs N+1 vs IN-batch across RTT, result-set size and arrival rate. MySQL 8, Hibernate 6, tc netem, k6.
- director-pattern A layer that holds your stance between you and your products, and why that is not a dispatcher.
- ccpulse Desktop tool that keeps a Claude Code token window aligned with a schedule. macOS, Linux, Windows.
Independent researcher. ORCID 0009-0004-4987-3897.
- When Does N+1 Break a System? A Preregistered Measurement of Query-Pattern Performance-Degradation Boundaries. First author. 10.5281/zenodo.21594741
- When Do Agent Loops Mistake Stagnation for Progress? Self-Evaluation Bias and Externally Grounded Verification in Long-Running Autonomous LLM Agent Loops. Co-authored. 10.5281/zenodo.21594735