Teaching an AI agent your domain's rules by hand works for a few — but does it survive hundreds?

This explores whether hand-writing domain rules into an agent's harness keeps working as the rules, domains and tasks pile up into a library. The corpus has no note that tests this directly, so what follows is pieced together from neighbouring work. The pieces point one way: codified rules pay off where they are stable and checkable, but hand-curating and hand-wiring them is the part that stops scaling.

The closest evidence is a study that ran automated optimization loops across many environments. It found four harness mechanisms (action execution, context compaction, observation handling and delegated reading) that cut token traffic by roughly 45–49% on a 51-task benchmark at comparable performance, and the gains were independent of model improvements (Can agent harnesses be automatically optimized across many environments?). The mechanisms that held up across that many environments were general plumbing found by search, not domain rules someone wrote down. That framing also gives a fair test. Multi-agent performance turns out to be largely a token-spending function, with about 80% of the variance coming from budget (How does test-time scaling work at the agent level?). A large rule library that mostly adds prompt text may just be buying tokens, so its benefit should be judged at equal cost.

Codifying does have a strong form. Code is executable, inspectable and stateful, so a rule written as code can be run and checked, while a rule written as prose can only be hoped for (Can code serve as the operational substrate for agent reasoning?). Where rules are precise and repetitive, small models can handle the resulting subtasks at 10–30× lower cost than large ones (Can small language models handle most agent tasks?). A verified rule layer can therefore be both trustworthy and cheap to run.

The strain shows up in three places. First, a rule library is capped by its author's imagination. Agents trained on expert demonstrations are limited to what curators imagined (Can agents learn beyond what their training data shows?), and hand-written rules plausibly inherit the same ceiling. That is an extrapolation, not something the note tests. Systems that generate a workflow per query, rather than reusing fixed task-level templates, are one response to it (Can AI systems design unique multi-agent workflows per individual query?). Second, finding the right rule becomes the bottleneck. Semantic capability vectors make discovery scale sub-linearly as the pool grows, and manual wiring does not (Can semantic capability vectors replace manual agent routing?). Third, the natural fix of splitting domains across specialized agents (Do single agents always hit organizational limits?) brings its own cost. Coordination degrades predictably with network size, and agents accept neighbours' information without checking it (Why do multi-agent systems fail to coordinate at scale?). A wrong rule passed between agents can spread rather than get caught.

The corpus suggests a practical rule of thumb. Codify what is stable and verifiable, retrieve rules instead of wiring them in by hand, and let automated search find the general mechanisms. The corpus has no direct evidence on whether a hand-built library of domain-specific rules holds up at scale.