Field note / 002Agents Need a Runtime, Not Just a Loop
Why durable state, resumable tasks, idempotent tools, human checkpoints, and evaluation are becoming the real architecture of agentic systems.
Read the essayNotes from the hard part.
Clear thinking on AI systems after the demo: memory, evaluation, retrieval, agents, and the engineering choices that make them useful.
Field note / 002Earlier field note / 001
The journal so far.
Agent Memory Needs Evaluation, Not More Context
Why capable AI agents need selective memory, observable retrieval, and evaluation across the full task journey.
Read field note01 / Evidence
Claims need a trace.
Architecture, constraints, and evaluation matter more than impressive vocabulary.
02 / Practice
Build before certainty.
Useful ideas get stronger when they meet real data, tools, latency, and users.
03 / Clarity
Short words. Deep work.
Complex systems deserve precise explanations, not longer headings.