Context Engineering

A list of my notes from Context Engineering for AI Agents: Lessons from Building Manus:

Cache prompt prefixes

• Keep prefixes stable (e.g. avoid dynamic prefixes like {timestamp})
• Append-only (validate stable key ordering when serializing JSON)

Dynamic tool loading

• Note: at time of writing, giving the agent more tools or loading tools on-demand worsens agent performance
• It's hard to manage tool hallucinations and schema violations as tool definitions bypass context window
• Manus uses "masking" instead (avoid changing/removing tool definitions, just specify constraints on subset of tools)

Context compression

• Inevitable information loss
• Mitigate by using File System as context/memory
• "Restorable compression strategy"

Recitation

• Manipulate attention by rewriting (having the LLM recite) objectives at the end of the context

Error recovery

• Avoid omitting failed responses/actions
• Use failure-as-feedback to improve future responses

Few-shotting

• Brittle formatting for few-shot examples can lead to pattern imitation that causes "drift, over-generalization, and hallucination"
• Address through controlled randomness (small amounts of structured variation/minor noise)