From the workbench
Practical writing on AI strategy, cleanup, and delivery from the Moonlight AI team.
Adding agents, tools, and autonomy doesn't make AI systems reliable. Here's what the evidence shows, and the ten design controls that make agents dependable.
How to design multi-agent workflows as graphs: remove waits that move no data, give nodes schemas, verify against anchors, and know when one agent wins.
An eight-step method for designing AI agent evals: build tasks from real failures, pick graders, report pass^k, calibrate LLM judges, and gate releases in CI.
Most AI work doesn't need a frontier model. Find each workload's threshold with evals, route to the cheapest model that passes, and know when not to.
GitHub was degraded for 7h47m on August 17, 2026. A scaling gap started it; retries kept it going. What it means for your CI, your SLA, and your AI clients.