Blog
Notes on shipping AI agents in production
Engineering deep-dives and product updates from the Arclight team.
EngineeringMay 14, 2026·6 min read
Why agent traceability matters more than agent intelligence
A smarter model doesn't help if you can't see why it made the call it made. Here's why we bet on observability first.
Read moreProductApril 22, 2026·5 min read
Policy guardrails vs. prompt engineering: pick both, in that order
Prompting an agent to 'never do X' is not a safety control. Here's how enforcement at the orchestration layer changes the math.
Read moreEngineeringMarch 30, 2026·7 min read
What actually happens when you roll back an agent in production
Rollback for agents is trickier than for regular code deploys. A walkthrough of what needs to be versioned and why.
Read moreCase studyFebruary 11, 2026·4 min read
What multi-model routing actually saved us (with real numbers)
We routed classification steps to a smaller model and reasoning steps to a frontier one. The cost impact surprised even us.
Read more