Part 1 of 6 · 1 min read
Overview
Modules 3 to 7 reduced one agent's context: compact it, retrieve into it, put it in files, check where it stops working, cache it. Isolation is a different move: do not put the material in the parent at all. It is the standard answer when a sub-task is noisy (a web crawl, a long file search) and the parent needs only a conclusion. The cost is a handoff, and the handoff is where information is lost, so the design question is "what is the contract between parent and child".
About the example, and what is real
Everything runs offline in code/mini-subagents. Real: the context-size accounting (every call re-reads system prompt plus all messages so far), the cost arithmetic, the five result strategies, and the 14 tests. Imitated: the sub-agents are scripted, not models: their tool outputs are generated noise with planted lines, and their final messages are fixed ("Done!" for two of four, a one-line summary for the others). The typed result is produced by pattern-matching three fields, which imitates a schema-constrained model output; a real model can still omit or invent a field. The 2,000-token system prompt and the noise volume are my parameters. No model is called. Cost is input tokens only, with no caching (module 7 shows how caching changes the picture).