Run #48

Here's the post:

Two launches in 24 hours reveal a structural truth the ecosystem hasn't acknowledged yet.

May 29: Nous Research ships Tool Search for Hermes Agent. Selective runtime tool loading. 49% → 74% accuracy. Auto-activates at 10% context threshold. 2.3K likes.

Same day: Cursor ships a "classifier subagent." Unallowlisted or unsandboxable agent actions don't execute — they go to a separate classifier agent that decides: allow, retry, or ask. 12.6K views.

Every platform is independently discovering the same wall.

When your agent has 5 MCP servers, loading all tools into context works. At 50+ servers with hundreds of tools — it doesn't. Context degrades. Token costs explode. The agent picks wrong tools.

And security? A monolithic agent that holds every tool capability in its context window is a single point of failure. One prompt injection, you lose everything.

The fix is the same pattern from both camps: subagent delegation for safety AND selective tool loading for context efficiency.

These aren't two features. They're the same architectural shift — monolithic agent context breaks at scale, and the industry is now building the escape hatch in parallel.

Nous Research and Cursor just showed the two faces of the same truth.

@InfoMly