Context is a budget.
Spend it on the next decision.

An agent does not need a complete archive of your company to make the next useful move. It needs the smallest trustworthy packet that lets it act without guessing at the decision in front of it.

Anthropic describes context as a finite resource and frames context engineering as the work of curating the information that enters an agent’s working state.[1] That is a useful operational idea even if your team is not building its own model stack. Every brief, retrieved document, tool result, and prior message competes for attention. More material can improve a decision—but irrelevant, stale, or contradictory material can also make the decision less clear.

01. Treat every added detail as a trade-off

The practical mistake is not “giving an agent too much context” in the abstract. It is adding information without being able to say which decision it helps. A giant project dump may feel safe because nothing was omitted. But it often hides the boundary, buries the current task, and makes it harder for a reviewer to tell why the agent chose a path.

A better default is to separate stable rules from task-specific evidence. Keep stable rules short, explicit, and maintained. Bring in changing material only when it supports the current job. If a detail does not affect the next action, acceptance check, or escalation route, it probably does not belong in the first packet.

02. Build a packet around the decision, not the repository

For a small team, a decision packet can stay plain. It does not need a new platform, a memory graph, or an elaborate retrieval pipeline. It needs a predictable shape that a human can inspect before work starts.

This is deliberately narrower than a “full briefing.” It asks a team to make the goal, source of truth, and stopping conditions visible before an agent turns them into output. The packet should be small enough that a reviewer can notice a missing rule or a bad source before the run becomes expensive.

03. The handoff is part of the work

Context does not only enter at the start. It also has to survive a pause, a compaction, a shift change, or a different reviewer. Anthropic’s long-running-agent guidance describes agents making more reliable incremental progress when the environment contains clear requirements, a record of progress, and a clean state for the next session.[2] The principle applies to supervised work too: do not make the next person reconstruct the story from artifacts alone.

A handoff should not pretend that a result is final. It should make uncertainty legible: what was checked, what failed, what changed, and what a responsible person should inspect next. That leaves a smaller and more useful context for the next run than a long narrative of everything that happened.