Blog / Field Notes
Periodic commentary on what is moving in agentic systems. A companion to Architecting Agentic Systems, time-bound where the book is stable.
2026
Don't leave the agent unbounded
August 28, 2026
August 2 applied Article 50 transparency. High-risk oversight remains December 2, 2027. IBM's production platform and an unattended Hermes run show the envelope has to sit in the dispatch path: identity at every hop, evals before the change, an approval prompt the operator cannot disable.
July 18, 2026
The mid-2026 conversation has moved from whether to build agents to how to wire them together, and the recurring answer is that the hard parts live in the seams: the loop, the tool-call boundary, the trust between agents, the memory write path. The field is rediscovering the harness.
July 5, 2026
What practitioners are saying about agentic systems in July 2026: reflexive agentification is losing ground to disciplined workflows, debugging is the open unsolved problem, and the three-placement oversight model is converging from practice and regulation as the EU AI Act deadline arrives.