What the labs say about context

Claim

Four production teams agree on the same rules. Find the smallest set of high-signal tokens. Keep the prompt prefix stable and append-only so the cache hits. Provide less up front and let the agent fetch. Use the filesystem as restorable memory. Name the failure modes: poisoning, distraction, confusion, clash.

Evidence

• Anthropic (2025-09-29): context engineering is "curating and maintaining the best set of tokens during inference"; "context rot" is a gradient; system prompts at the "right altitude"; tools with minimal overlap; for long horizons use compaction, structured note-taking and sub-agents; "do the simplest thing that works"; smarter models need less prescriptive engineering.
• Manus (2025-07-18): KV-cache hit rate is "the single most important metric"; about 100 to 1 input-to-output ratio makes prefill dominant; cached input is 10x cheaper. Rules: no timestamps at the top, append-only, deterministic serialisation, mask tools rather than remove them, filesystem as context ("restorable compression"), a recited todo file, keep failures in context, avoid few-shot pattern drift.
• Cursor (2026-01-06): dynamic context discovery; long tool outputs to files, chat history as a searchable file, skills and MCP descriptions loaded by name; A/B showed a 46.9% reduction in total agent tokens on runs that called an MCP tool.
• Claude Code team (2026-04-30): the team alerts on cache hit rate and declares incidents when it drops; layout is static system prompt and tools, then CLAUDE.md, then session, then messages; past breakages were a timestamp, non-deterministic tool ordering and changed tool parameters; never swap models or tools mid-session; compaction is a cache-safe fork producing about 20k-token summaries; model switches via subagent handoff.
• LangChain: write, select, compress, isolate taxonomy.
• Breunig (2025-06-22): poisoning (a hallucination gets re-referenced), distraction (over-reliance on long context past about 100k tokens), confusion (every model worse with more than one tool on the function-calling leaderboard), clash (sharded multi-turn prompts cost 39% on average).

Open questions

• Whether Manus's logit-masking advice survives now that providers offer allowed_tools style restriction natively.

Sources

• https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents — blog, Sun Sep 28
• https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus — blog, Thu Jul 17
• https://cursor.com/blog/dynamic-context-discovery — blog, Mon Jan 05
• https://claude.dev/blog/lessons-from-building-claude-code-prompt-caching-is-everything/ — blog, Wed Apr 29
• https://www.langchain.com/blog/context-engineering-for-agents — blog, Tue Jul 01
• https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html — blog, Sat Jun 21

AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.