Why better AI agents sometimes need fewer tools

By Pascal Bouman··3 min read
Product team reviewing a simplified AI agent workflow with fewer tools.

The choice: assess existing tools first

Retain a tool only when it demonstrably makes a unique contribution to a clearly defined task. This is not a general rule that less is always better, but a practical choice when tools overlap or their function is unclear. The ToolScope passage describes how real-world toolsets may contain redundant tools with overlapping names and descriptions; this overlap introduces ambiguity and reduces selection accuracy. The same passage mentions strict input-context limits, meaning large toolsets cannot always be considered efficiently. Therefore, identify for each task which tool the agent genuinely needs to be able to select, which path is duplicated, and what information makes the tool description distinctive.

Measure the entire path, not just the end state

A successful final task is not enough to justify a tool. The benchmark passage on agent trajectories states that existing evaluations primarily look at task success and can miss execution efficiency. It also describes redundant steps that consume many resources while contributing little to task completion. Record the following for each test: task type, selected tool, number of steps, error or recovery, human correction, and final outcome. Then compare a path with the tool against one without it or with a merged variant. A lower step count is not, by itself, evidence of quality; keep the task the same and check whether the outcome is still usable.

Make removal a manageable decision

Use a short decision log for every tool you retain, merge, or pause: task, unique function, owner, error handling, measurement window, and follow-up decision. For each tool, note the specific error a person catches and when the path will be reviewed. Practical aid: Record for each tool: task, unique contribution, owner, error handling, measurement window, and decision (retain, merge, or pause). The available passages concern overlap, selection, context, and redundant steps; they do not prove that a specific tool should be removed from your agent or what the optimal toolset size is. The available sources do not support a universally optimal toolset size or a guarantee of better task completion for an individual agent.

Audit table assessing AI agent tools for necessity and risk.
Your personal AI research team

Developments move too fast to keep up with everything yourself.

You need a research team that tracks changes, checks sources and decides what matters for your work.

Choose what you want to follow and receive only the updates that matter to you.

Updates tailored to your interests
Researched by specialist agents
Relevant insights, not daily noise

What do you want to follow?

You receive a confirmation email first and only join after clicking it. See the privacy policy.

Latest articles

Recent knowledge base articles selected for this page.