Get the Cheat Code on Long-Running AI Agents—Here’s What Manus, Google, and Anthropic Learned After Trial and Error + 12 Prompts to Help Build Long-Running Agents Yourself

Source: Nate’s Newsletter Date: 2025-12-09 URL: https://natesnewsletter.substack.com/p/i-read-everything-google-anthropic

Summary

Nate synthesizes what Manus, Google, and Anthropic independently discovered about long-running AI agents: larger context windows don’t solve performance degradation — the real problem is attention dilution, where critical information gets buried in accumulated tokens. The solution isn’t more context but “compiled context” — architectures that actively manage what the agent sees at each decision step, separating working context from stored memory across a four-layer model.

Implications

• Agent-product positioning thread. Context engineering is now a first-class architectural concern, not a prompt-writing trick. Products and teams that treat long-running agents as naive context accumulators will see the same 10-20 minute performance cliff these three organizations documented — it’s a design pattern, not a model quality issue.
• Enterprise adoption thread. The convergence across Manus, Google, and Anthropic signals this is a fundamental architectural requirement rather than a temporary limitation. Enterprise agent deployments need to budget for context management infrastructure, not just model access.
• Watch: Whether the four-layer memory model (working context, sessions, memory storage, artifacts) becomes a standard reference architecture for production agents, and which frameworks implement it most cleanly.