Hermes v0.15 Tool Search: Agent Accuracy Jumps from 49% to 74%

Key Takeaways

• Tool search activates automatically when your loaded tools fill more than 10% of the agent's context window — no configuration required.
• Tool-selection accuracy improved from 49% to 74% in Nous Research internal benchmarks; the agent picks the correct tool roughly 25 percentage points more often.
• Large toolsets (100+ tools) are now practical: the agent stays sharp because it only loads what it needs for the current task.
• v0.15 also splits the agent into a planner role and a doer role, each maintaining its own memory and skill library that grows over time.
• Kanban-style swarms let multiple agents coordinate parallel tasks on a shared board, handing work off as each step completes.
• Update with a single terminal command: hermes update.

The Context Window Problem Tool Search Solves

Every tool you load into an agent consumes context window space — the agent's working memory. Load 50 tools at startup and the agent spends a significant chunk of its thinking capacity just holding the tool descriptions, leaving less room for actual reasoning. This is why heavily tooled agents have historically been slower and less accurate: they're cognitively overloaded before the first task begins.

Tool search inverts this. The agent keeps a handful of always-available tools (web search, file I/O) and maintains a searchable index of everything else. When a task arrives, it queries the index, pulls in the relevant tool, uses it, and moves on. The context window stays clear for reasoning rather than being cluttered with tool definitions the current task doesn't need.

Planner/Doer Split

The v0.15 update also introduces a two-agent architecture for complex tasks. A planner agent decomposes the work and assigns subtasks; a doer agent executes them. Both agents build their own skill libraries independently — the planner gets better at decomposition, the doer gets better at execution — rather than sharing a single model that has to be good at both. This mirrors how effective human teams work: a project manager and an implementer with distinct expertise.

Commands & Code Mentioned

hermes update

Related on OpenClawDatabase

• Hermes Agent Hub — overview, setup, and all guides
• Long-Running Tasks & Scheduling — swarms and parallel execution
• Hermes MCP Tools Guide — connecting tools with the new catalog
• Compare agent platforms — Hermes vs OpenClaw vs Claude Cowork