Agent Status String

Agent Status String – a summary of dynamic metadata inserted at the end of the context: tool call counters, TODO progress, current time, environment state. Analogy – phone status bar.

Theoretical basis: in-context learning is more like retrieval than reasoning – the model is good at finding, but bad at aggregating itself.

Requirements:

the status string should be maintained by code, not LLM;
it should cover all possible questions – otherwise it becomes a “false authority”;
its accuracy is a first-order production metric.

Effect: saves strong models 80–90% of thinking tokens, raises accuracy of weak models by tens of points.