The uncomfortable truth: more tools make most AI agents worse
AI agent tool optimization is the practice of limiting an agent to the smallest, most relevant set of tools and contextual data needed for the current task so that its limited context window is spent on work that matters instead of on unused integrations, bloated instructions, or redundant references that quietly drain performance.
Connecting every integration you can find feels powerful, but it is one of the fastest ways to wreck coding agent performance. One developer wired in MCP servers for design apps, note-taking tools, meeting recorders, cloud services, and more until their agent had 172 tools available; with Tool Search turned off, those 172 tool definitions ate 141k tokens before they even typed a word. When most of a context window is gone on startup, the agent fails at basic tasks not because the model is weak, but because you left it no room to think. If your agent feels flaky, the problem is probably not “it needs more tools” but “you have already given it too many.”
Why overloaded tools quietly destroy coding agent performance
Every AI agent lives inside a context window, a fixed budget of tokens that must hold system prompts, conversation history, working code, and the descriptions of every tool the model can call. When you bolt on server after server, those tool descriptions swell until they consume the same memory the agent needs for your repository. In the 172‑tool setup, 141k tokens vanished into definitions alone, long before any code or instructions were loaded.
That waste has two ugly effects. First, every token spent describing a tool is a token that cannot hold your files, so the model runs out of room, loses track of earlier context, and chokes on exactly the long, multi-file jobs where you expect it to shine. Second, the more tools you present at once, the more the model’s job shifts from solving problems to picking from a crowded menu: it must choose between dozens of similar tools and can stall or call the wrong one when names overlap. Overloaded configurations do not make agents smarter; they turn them into indecisive switchboards.
Hermes-style workflows: cutting waste without losing power
Hermes Agent-style workflows prove you can slim down tool exposure and still increase workflow automation efficiency. The core idea is architectural: keep your data and skills on infrastructure you control, and send the model only the information it needs. This approach uses three pieces working together: a local context store for all data the agent touches, a skill library of reusable guides and reference material, and a minimal-prompt extractor that pulls only relevant slices into the model call.
Here is how that looks in practice. Fresh data from systems like a CRM or data warehouse lands in the context store; when a task arrives, the extractor finds the most relevant information, combines it with matching entries from the skill library, and sends only that bundle to the LLM. The result comes back and is stored again, so every task enriches the next one. This context-centric pattern means the model never sees entire knowledge bases or redundant tool definitions; it sees the minimum needed to act. You cut token waste, keep capabilities, and get agents that stay coherent across long-running work.
From tool hoarding to tool selection strategy
The practical fix is not “no tools” but disciplined tool selection strategy. One coding agent user moved away from the everything-connected mess by checking which MCP servers were active and turning off anything they would not use in the current session. For integrations they wanted disabled by default, they added those servers to a blocklist, while a small handful of always-needed servers were marked to load on every turn.
Tool Search, a feature that defers tool definitions until the agent needs them, shows the impact of this approach on AI agent tool optimization: with Tool Search off, 172 tools loaded upfront and consumed 141k tokens; with it on, the same 172 tools initially recorded as zero tokens because their definitions were deferred until needed. That is the pattern to copy manually if your stack does not have Tool Search: start with a minimal set, run real tasks, and add tools only when they directly improve outcomes, especially for recurring procedures rather than one-off convenience.
Designing workflows that compound instead of reset
Prompt-centric habits keep you stuck in a loop: you type a request, the model responds, and everything resets on the next turn. Hermes-style, context-centric workflows flip this by treating every run as another layer of shared memory. The context store accumulates everything the agent touches so that each task contributes to the next one, like a brand brief written in January later serving as reference for a campaign brief in June without hunting through old chats.
The same pattern applies to teams. When a new coordinator joins midway through a project, they should not need to scroll back through weeks of chat logs to learn the brand voice, audience definitions, or competitive positioning; those live in the skill library and context store, available to both the agent and the human from day one. The second data pipeline in this model uses scheduled jobs and a relevance filter, but the core remains “store first, extract, then call the model” for every batch. If you want better coding agent performance and workflow automation efficiency, stop wiring every possible tool into the model and start wiring knowledge into your context instead.
