LLMs, GenAI, RAG, agents, evaluation · insight-1557
Compact long agent runs at task boundaries, keeping decisions and open items, not raw tool output
- Evidence type
- documented fact
- Difficulty
- medium
- Roles
- Applied AI, ML engineer
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
An agent run approaches the context limit after many large tool results, and quality begins to drop.
Action
Summarise at natural boundaries, keeping goals, decisions made, open issues and references to where details live, clear stale tool results, and keep durable notes outside the window to re-read.
Reason
Context engineering guidance describes compaction that keeps decisions and open bugs, clearing old tool results as the gentlest form, and notes stored outside the window.
Boundary
Agents with long, tool-heavy runs.
Symptom or risk
Old raw outputs crowd out the instructions and decisions that matter, or an automatic cut discards the one fact still needed.
Exceptions
Short runs do not need it, and summaries can drop details; keep the originals retrievable.
Trade-offs
Summaries cost a model call and may lose information.
Pitfalls
Summarising away the exact identifiers or error messages that later steps need.
Lessons and concepts
Claims and their sources
Anthropic's context engineering guidance describes compaction that preserves key decisions and open issues, tool result clearing of old raw outputs, and structured notes kept outside the context window.