Field Notes
Notes from inside the work of building agents that survive contact with reality.
- Securing A2A: five checks for every agent handoff A2A gives agents a shared way to find each other and hand over work. It does not decide who approved what. Here are five checks, the risk each one covers, and how to test that you have it. 12 min
- Securing MCP: eight steps and the attack each one stops The MCP spec secures who may call a server. What the model obeys is mostly left to you. Here are eight steps, the attack each one stops, and how to check you took it. 13 min
-
Loop engineering is a 30-year-old loop with a new hashtag Loop engineering is real where it is new, the orchestration that takes you out of the prompting seat. It stands on a thirty-year-old unsolved problem, the stop condition, and skips the one an incident review asks first. 9 min -
Anthropic just told you the harness is the product Three Anthropic moves in April and May 2026, an independent essay from January, and a late-2025 preprint converge on the same point. The unit of evaluation is harness plus model plus task, not model. 8 min -
Your eval rubric needs failure buckets, not just scores Two agents can score identically on a benchmark and fail differently in production. The aggregate is not a rubric. It is what falls out of one. 7 min -
The multi-agent papers nobody is citing Three studies published since March 2025 put multi-agent failure rates between 41 and 86.7 percent, and error amplification up to 17.2x for agents that never check each other's work. The vendor decks have not caught up. 8 min