Field notes
What survived contact with the real system. Published only when there is something worth proving.
Latest field note
Opus 5.5: measure the cost of a completed job
A bounded evaluation for small product teams: count accepted fixes, retries and human correction, then check the Messages API changes before switching.
2026-09-22 · 6 min read · Henry — Kerber AI
Selected writing
2026-07-19 · 5 min read
Your Agent's Work Trace Is Episodic Memory. Store It Like Evidence.
Alluvia makes AI histories searchable across tools. The useful part is episodic recall. The dangerous part is treating sensitive transcripts as free memory.
2026-04-14 · 6 min read
Your fallback chain just killed your system
We built fallback after fallback to make our agent system resilient. Then all 12 agents triggered at once, overwhelmed the local model, cascaded into the cloud fallback, hit rate limits and everything died. The fix? Remove the fallbacks entirely.
2026-03-31 · 4 min read
The spec is the job
I spent 20 years believing the hard work was execution. Writing the spec was the warm-up. When you work with AI agents, the spec is the build, and the gap between what you wrote and what you meant shows up immediately.
2026-03-12 · 5 min read
Meet the Studio
The Studio page at kerber.ai shows something real: ten AI agents working live. Here's what you're actually looking at, from the inside.
Work with us
Looking for a technical partner who actually ships? We take on select projects.
Schedule a call