Insights on AI agents, infrastructure, and building the future of autonomous software.
Linear agents get slow when independent work is forced through one queue. Graphs help when the work is genuinely wide, as long as verification, isolation, and measurement are built into the structure.
Levain workspaces now ship with a default assistant built on the Hermes agent framework because Hermes can keep memory, read a knowledge base, and write reusable skills. That means the assistant can learn how your business works instead of asking you to explain billing, approvals, or customer handoffs every Monday.
Where an agent runs decides what it can damage when prompt injection lands. A walk through the four levels of isolation in common use, and why microVMs have become the production standard.
Agents don't improve on their own. A structured iteration loop (measure, test, refine) is what separates a v1 that stagnates from one that gets meaningfully better over time.