Less Context, Better Agents: Efficient Context Engineering for Long-Horizon Tool-Using LLM Agents

An agent fills in an expense form, reads the updated balance and moves to the next item. Its conversation now contains several versions of the same form. Keeping every response preserves history, but also gives the model many outdated balances and control states to interpret.

This Microsoft study evaluates a simple context policy in Dynamics 365 Finance and Operations: retain recent tool interactions and summarize selected older ones. On the primary hotel-expense benchmark, pruning plus summarization improves the reported complete-itemization rate while reducing total tokens and wall-clock runtime relative to full history.

The comparison has an important boundary. The main configurations all include a separate user model that answers the agent's follow-up questions. A much weaker baseline omits that participant and often stalls. Its results cannot be used to attribute the entire improvement to context engineering.

The lesson is practical and narrow: a bounded working context can outperform a transcript containing stale operational state. The right window depends on the workflow, and a generated summary should never replace an authoritative read of balances, permissions or completed actions.

Enterprise tools often return more information than the next decision needs: form metadata, navigation history, repeated records and full state snapshots. An agent can spend most of its inference budget rereading that material while still missing the latest relevant change.

This is a useful setting for context-window management. Measure whether the agent sees current state, retains unresolved requirements and knows when a write has actually succeeded. A shorter prompt is valuable only if those properties survive.

The paper also encourages an independent definition of completion. Reading the saved application state is stronger than trusting “done” in an assistant message. For a real financial workflow, extend that check to line identities, categories and exact amounts; a zero remaining total alone cannot establish that every item was classified correctly.