What should an agent remember between steps?
A paper-inspired look at whether structured state and the latest observation can replace a growing conversation history without losing a host-verified result.
Planned follow-upField notes
Notes
Engineering field notes on the boundaries, evidence, and operating decisions behind practical AI systems.
Series · Engineering a Reusable Custom Agent Loop
One controlled experiment. One reusable agent loop.
Start with the recorded showcase: a generic agent loads its task, works inside an isolated boundary, and reaches host-verified storage. The following articles will unpack the problems and engineering decisions behind that loop.
Engineering the Agent Loop
Watch GPT-5.6 Terra perform task-specific work through a harness built on Google ADK and designed to be model-agnostic, while the host validates the result, persists accepted records, and verifies stored state.
Watch the showcase → Personal companionLessons in building an autonomous agent loop
Why I built a custom loop to experiment with context and delegation, and what I learned getting it to complete a task.
Read the field note →Why the host owns “done”
Sandboxing generated code, deterministic acceptance, persistence gates, and a fresh reread of stored state.
Planned follow-upHow the solution layer evolves across tasks
How accepted approaches become versioned skills and adapters—and the evaluation needed before future agents inherit those updates.
Planned follow-upAlso planned: a deeper look at hidden acceptance contracts and why an earlier run stopped when its required fields were not clear to the agent.
Other field notes
Additional ideas remain in drafting.
When should an AI system act instead of just classify?
A practical view of the jump from perception to decision and from decision to operational action.
Drafting nowWhat sports tracking teaches about real-world AI systems
Tracking players is only one part of the challenge. Reliable systems need context, temporal reasoning, and usable outputs.
Drafting nowHow to think about AI decisions in telecom and infrastructure
Spatial optimization and visual evidence create a useful lens for designing AI systems that work beyond the screen.
Drafting nowNext step
Want the notes when they publish?
The use-case pages are live now. Join the waitlist for new essays and practical notes as they publish.