# The agent said it saved a file. The file tool was never called. How do you catch this? — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/4a09cbd6443d809b

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-10-14 03:46 ingested 2026-10-15 02:37

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1x5h918/the_agent_said_it_saved_a_file_the_file_tool_was/)

## retained text



```
I'm building a local agent for Windows, and I'd like to hear how others handle this.

I had a small local model (4B) research something, then asked: "Give this to me as an HWP or Word file." Five seconds later it replied with a save path and "Saved."

I opened that path. Nothing was there. The file-writing tool had never been called. The agent reported saving to a location that didn't exist.

Two things caused it.

1. The router didn't read "give it to me as a file" as a request to create a file. It knew "make" and "save," but not "give," "pass," or "send." So the request went down the path that answers in words only, with no tools.

2. On that path, nothing checked whether the answer claimed an action that had actually happened. A model with no tools can only write plausible sentences.

What I changed:

\- A request shaped like "\[format name\] + as + give/pass/send" now counts as a file request. I matched the shape instead of adding verbs one at a time. Before shipping, I ran the old and new rules side by side over 1,931 real requests: the new rule caught 2 more, and neither was wrong. An earlier, looser version was dropped because it also matched "by keyword" and "answer in Korean."

\- Before a no-tool answer goes out, if it contains words like "saved" or "created" together with a file name or path, the agent checks whether that file exists. If it doesn't, the answer is discarded and the request is sent back through the tool path.

After the fix, the same request produced a real file with the content in it. I only ran it a few times, though, and in one run it stopped to ask "HWP or Word?" instead of finishing.

The part that bothers me is that this isn't the first time. I've seen the same pattern twice before:

\- It replied "I checked" without calling a single tool.

\- 1 of 9 direct answers was not new at all. It was a copy of an earlier answer in the same conversation.

All three are the same mismatch: what the agent says it did versus what it actually did. Each time I blocked that one type. "Emailed," "scheduled," and "deleted" look like the next ones.

**Questions**:

1. Where do you catch the agent's "I did it"? Before the answer goes out, or does a person notice later?

2. Do you write a separate check per type of claim, or is there one rule that covers them, something like "block any claimed action that has no matching tool record"? I'm still blocking one type at a time.

3. Does this happen more with smaller models for you? It did for us.

4. Has your verification ever been strict enough to throw away good answers? How did you find the line?
```



document fetched 2026-10-15 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1x5h918/the_agent_said_it_saved_a_file_the_file_tool_was
