The practical version of this question is not only “how much context can we compress?” but “which stable worker state should be assembled once before inference?” Paper 3 reports a bounded study of repeated state serialization versus pre-inference deterministic worker construction. It is a useful way to separate task-relevant context from repeatedly rebuilding role, authority, constraints, and history.
How much of AI spending is paying a model to reread its job description?
A bounded study, reported as Paper 3, compares repeated state serialization against pre-inference deterministic worker construction in AI systems, framing the question as which stable worker state should be assembled once before inference rather than only how much context can be compressed. The study separates task-relevant context from repeatedly rebuilding role, authority, constraints, and history. The source provides no figures, named authors, or quantitative results.
Build each agent's stable role, authority, constraints and history deterministically once before inference, then pass only task-relevant context per call instead of re-serializing full state.
more in playbook: Designing agent systemsRun your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.