Context Engineering for Agents That Work for Hours
How to keep long-running agents coherent with compact state, durable artifacts, explicit plans, and resumable execution boundaries.
As agents move from one answer to long sequences of work, raw conversation history stops being a reliable memory system. Logs grow, instructions compete, and crucial decisions disappear inside tool output. Context engineering is the discipline of deciding what the model needs now, what belongs in durable state, and what can be safely summarized.
Separate the working set from the record
Keep the immediate objective, constraints, current plan, and latest observations in the active context. Store files, source material, completed results, and audit evidence outside it. The agent should be able to retrieve the right artifact without repeatedly carrying the entire history.
- A concise objective and definition of done
- An explicit plan with completed and pending steps
- Durable artifacts for outputs that must survive compaction
- A decision log for assumptions, approvals, and irreversible actions
- Checkpoints that let the workflow resume without reconstructing hidden reasoning
Compact facts, not uncertainty
A useful summary preserves verified facts, unresolved questions, constraints, identifiers, and next actions. It should not turn a tentative inference into a settled decision. Compaction is successful when a fresh worker can continue from the summary and the artifacts without repeating finished work.
Design for interruption from day one
Long-running agents will hit time limits, unavailable tools, changed data, and human approval gates. Treat each boundary as a normal state transition. Resumability is not an operational extra; it is a core product behavior for dependable agent systems.
Primary sources
First-party documentation and announcements used to ground this field note.
