Recommended Free Tools
In one small experiment, adding a sentence about when to use a custom MCP tool to its description was followed by agents calling that tool, and no calls were made when the sentence was absent. The author’s 24-run comparison recorded seven tool calls in the runs with the sentence and zero in the runs without it. That is a narrow, context-specific result from three AI coding agents and one custom tool. It is not a general rule about how agents choose tools, and the author says so.
What the experiment tested
The author built a custom MCP server tool that sends several independent tasks to subagents for parallel execution. The prompt given to the agents did not name the tool, so any call had to come from the agent’s own choice after reading the available tool list. The tool description was the variable under test.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
The comparison covered three agents: OpenHands, OpenCode, and Qwen Code. Each agent was run with and without a startup rule file, and each of those six combinations was run twice with the usage sentence in the tool description and twice without it. That design produces 24 runs in total, which is why the sample is small: each cell of the comparison rests on two runs per agent.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsThe results
The tool was called in 7 of the 12 runs that included the usage sentence and in none of the 12 runs that did not. Breaking the result down by rule-file presence shows where the calls happened:
#1 Best Overall
| Startup rule file | Description included when-to-use sentence | Description without the sentence |
|---|---|---|
| Present (6 runs) | Calls in 5 of 6 runs | Calls in 0 of 6 runs |
| Absent (6 runs) | Calls in 2 of 6 runs | Calls in 0 of 6 runs |
| Total (12 runs each) | 7 calls | 0 calls |
Two features of this table matter for interpretation. First, the sentence was the only description change, but the rule file was also present in most of the calling runs. Second, with no sentence, the rule file produced no calls in this set of runs. Those two facts are why the author stops short of attributing the calls to the sentence alone.
Calls are not completions
The author measured tool calls separately from whether the delegated work finished. Of the seven calls, four ended with subagents completing their tasks. The per-agent picture is uneven:
- Qwen Code: the article reports subagent tasks completing in two of three calls, and seven of eight in a third set of calls.
- OpenCode: one run completed after the author corrected an error in the launch script.
- OpenHands: made calls, but none of the 32 subagents it launched completed. The author attributes this to a defect in the measurement program rather than to the agent’s behavior.
Across the runs, all eight tasks passed their tests, and the agents did not rewrite the tests to make them pass. Because the OpenHands completions are missing for a measurement reason, treat the completion figures as a partial picture rather than a clean success rate.
A separate result for rule files
The author also ran a distinct experiment using a consultation tool. In that test, the tool was called in all six runs that included a consultation provision in the rule file and in none of the six runs without it. The tool and the task differ from the parallel-execution test, so the two results cannot be placed side by side as a single effect size.
Why the result does not isolate a cause
The author’s first explanation was that processing weight, meaning how heavy the agent’s workload was, decided whether the description or the rule file mattered. That interpretation was challenged in review and withdrawn. The author now treats it as a hypothesis that fits the pattern without proving it. In the author’s words: “I now write the results of the two measurements separately.”
Several differences between the experiments prevent a clean attribution:
- The parallel-execution and consultation tests used different tools.
- The tasks differed, and the parallel task involved several independent subtasks.
- The description guidance was not identical in quality across the experiments.
- The rule-file instructions varied in specificity.
A stronger follow-up would vary the description guidance and the rule-file instructions independently, keep the tool, task, and agent versions fixed, and record the outcome in the same way for every run. Without that design, the sentence can be treated as a plausible lever, not a demonstrated one.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Related studies are about a different question
The author cites two arXiv papers on tool descriptions. Both address whether better descriptions improve task success, not whether an agent decides to call a tool in the first place, so they should not be read as direct confirmation of the call-frequency finding.
Best Value
- “MCP Tool Descriptions Are Smelly! Towards Improving AI Agent Efficiency with Augmented MCP Tool Descriptions” (arXiv 2602.14878, submitted 2026-02-16, revised 2026-05-31). As summarized in the article, the study sampled 856 MCP tools across 103 MCP servers. It found that 97.1% of the sampled descriptions had at least one defect and that 56% did not state their purpose explicitly. Augmenting descriptions produced a median 5.85 percentage-point increase in task success, with execution steps rising 67.46% and performance declining in 16.67% of cases.
- “Learning to Rewrite Tool Descriptions for Reliable LLM-Agent Tool Use” (arXiv 2602.20426, submitted 2026-02-23, revised 2026-04-29). As summarized in the article, rewritten descriptions improved per-query success by an average of 60.89% over the original descriptions in an experiment with at least 150 candidate tools.
These figures come from the article’s account of the papers. Check the papers themselves before relying on the exact numbers.
What the specification says
According to the author, the MCP 2026-07-28 specification describes a tool’s description as a human-readable explanation of its functionality. It does not require the description to say when the tool should be used. A when-to-use sentence is therefore an addition beyond the specification’s minimum, not a conformance requirement.
Practical steps if you want to test this on your own tool
- Write the description as it stands, then add one sentence that says when the tool should be used and what kind of task it fits.
- Keep the tool, task, and agent version identical across the two versions.
- Decide in advance whether you will count calls, completions, or test passes, and record each one separately.
- Log the exact description text each agent received, so you can confirm which version a run used.
- Check your measurement program on a known-good run before reading results, since the author’s OpenHands completion counts were distorted by a logging defect.
- Vary the rule file and the description in separate runs, so that any effect can be assigned to one factor.
Published on September 29, 2026, the original article is the primary source for the experiment and its figures: Once I wrote in the MCP tool description when to use the tool, AI agents called it. The author states that the cited public materials were checked on September 10 and September 27, 2026.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
