Context management

What it is

Context management is the discipline of curating what information — instructions, tool results, prior conversation, retrieved documents — actually sits in a model's finite context window at each step, since a model can only reason well over what's currently in front of it.

Why it matters for you

Long-running agents accumulate tool outputs and history fast; once the context window fills with stale or irrelevant detail, response quality and reliability degrade even though the underlying model hasn't changed. If an agent that worked well on short tasks starts behaving erratically on longer ones, context bloat is the first thing to check.

Common failure mode

Naively appending every tool result and message to the context forever, rather than summarizing, pruning, or fetching-on-demand — the agent's context window fills with noise, relevant details get crowded out, and cost per call rises even as answer quality falls.

This is exactly what we work through in a session.