Engineer the Context, Not the Model

Manus chose context engineering over end‑to‑end training to iterate fast and stay model‑agnostic. They center on KV‑cache efficiency, stable append‑only contexts, action constraining via logits/prefill, file‑based memory, attention recitation, visible errors, and diversity to avoid few‑shot mimicry. The takeaway: agent success depends on how you shape context—cost, speed, recovery, and scale follow from it.

Key Points

• Optimize for KV‑cache hit rate: keep prompts stable, use append‑only, deterministically serialized contexts, and mark cache breakpoints; enable prefix caching and consistent session routing.
• Do not dynamically add/remove tools mid‑iteration; instead, mask logits and use response prefill plus a state machine to constrain action selection without breaking caches.
• Use the file system as externalized, restorable memory to handle large observations, control costs, and avoid long‑context degradation.
• Manipulate attention via recitation (e.g., a constantly updated todo list) to keep goals in the model’s recent focus and prevent drift.
• Keep errors in the context so the model can learn from failures; avoid over-uniform few-shot traces by injecting structured diversity to prevent behavioral lock‑in.

Sentiment

The discussion is uniformly supportive of the article. Commenters validate and extend its core ideas with particular enthusiasm for the file-system-as-memory concept and parallels to established software engineering practices. The only nuance comes from the economic incentive discussion around caching, which adds context rather than disagreement.

In Agreement

• File system as memory is validated through real-world practice: using an .agent/ directory with structured naming conventions eliminates the need for complex memory backends like MCP or RAG
• The simplicity-first approach is echoed by OpenAI Codex's own tooling design, which uses a straightforward update_plan mechanism
• Context engineering parallels software engineering best practices: don't let repos bloat, preserve error history via rollback PRs over deleted commits, and avoid repetitive patterns
• Fixed-price API providers have natural incentives to optimize KV-cache since it improves their margins, while metered plan users are just beginning to appreciate caching's importance