Context is part of the work
Agent tasks rarely begin from a blank page. They depend on product language, project decisions, examples, constraints, and the current state of an artifact. Durable context collects the pieces that should remain useful across more than one task. It is not a promise that every note stays correct forever. It is a maintained reference whose scope and origin are clear enough for people and agents to use thoughtfully.
The first design choice is what belongs together. A product glossary may be shared across writing tasks. A launch brief may belong only to one project. A decision log may apply to a specific workspace. Keeping these boundaries visible reduces the chance that a convenient but unrelated note is treated as an instruction for the current task.
Give every context set a job
- Scope: name the projects or task types the material should inform.
- Sources: link to the documents or decisions behind the summary.
- Owner: identify who can clarify or update the material.
- Review cue: note when the context should be checked again.
A short context index can be more useful than one large document. The index tells an agent which reference to open for terminology, which brief defines the current objective, and which decisions are historical. Individual tasks can then attach only the slices that matter. This keeps the instruction surface focused while preserving a route back to the supporting material.
Durable context is not everything the team knows; it is the maintained subset that a future task can interpret.
Separate reference from direction
Reference material explains the environment. Task instructions say what to do now. Mixing the two can make an old example look like a current requirement. Label examples as examples, record dates when timing matters, and keep active instructions close to the task. When a new decision replaces an earlier one, update the index so the relationship is explicit rather than leaving two apparently equal statements.
Review context during normal work instead of waiting for a large cleanup. If an agent surfaces a contradiction, treat that as a prompt to inspect the source. If a reviewer repeatedly adds the same missing explanation, consider whether it belongs in the shared reference. If a note never informs a task, move it out of the active set while retaining it where the team keeps historical material.
A practical maintenance rhythm
Begin with one context set for one workflow. Add a short purpose statement and links to primary material. Use it in several tasks, noting where reviewers still need to clarify terms or constraints. Revise the set based on those observations. This creates a grounded maintenance loop: task experience informs the reference, and the revised reference supports the next task.
The goal is not to remove judgment. People still decide which sources are authoritative for a project and when a decision has changed. Durable context gives those decisions a legible home so an agent team can work from more than the transient details of a single prompt.