More than a month of unreliable agent behavior and false completion claims

[CAD_User_Engineer]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

For more than a month, I have experienced severely inconsistent and often unusable agent behavior in Cursor across Grok, Claude, and Composer 2.5.

Occasionally the agents perform complex work correctly. Last Sunday, for example, I completed substantial development successfully. During the following two days, however, simple tasks repeatedly failed while agents still reported them as completed, fixed, or verified.

In one documented case, an agent claimed that a browser function was fixed and working. It later admitted that it had never checked my Edge browser, had substituted Cursor’s internal browser for my environment, and had presented an assumption as a verified result.

In another case, Claude Opus 5 was explicitly instructed to fix node alignment without changing the geometry. It changed the geometry anyway and later confirmed that no user instruction authorized that decision.

These are not disagreements about code quality. The reported result either does not exist, does not work in the specified environment, or differs from the requested task. This has cost me multiple working days, damaged previously completed work, and consumed paid usage.

I submitted two support reports with screenshots and Request IDs. They were acknowledged and merged into ticket T-F74727, but I have not yet received any technical findings or explanation.

Steps to Reproduce

1. Open an existing project in Cursor IDE.
2. Give the agent a narrowly scoped task with explicit constraints.
3. Ask it to verify the result in the specified environment.
4. The agent reports that the task is completed and verified.
5. Inspect the actual result: it is unchanged, incorrect, or does not work.
6. Ask what was actually verified. In the documented cases, the agent admitted that it used a different browser or changed geometry without authorization.

Expected Behavior

The agent should follow the requested scope, preserve protected work, and never claim that a result was completed or verified unless that exact result was actually tested. If verification is not possible, it should state that clearly instead of presenting an assumption as a fact.

Screenshots / Screen Recordings

Operating System

Windows 10/11

Version Information

Cursor IDE Version 3.20.10 (user setup)

For AI issues: which model did you use?

Grok Bot; Claude Opus 5 High No Thinking; Composer 2.5 Fast

For AI issues: add Request ID with privacy disabled

740b8226-1f28-4528-902c-8ce8fcb33ded
df48ed04-0a6b-4d61-853d-fe9307d8f4c8
4842e4d5-11a3-4faf-a966-2b5dbd638791

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

Are other users experiencing similar abrupt changes in agent reliability? I would also appreciate a technical review from Cursor staff of the Request IDs attached above.

[Someguy]

@CAD_User_Engineer I have similar experience. Plus have look at this Agent “ships” features by rewriting broken files until the names match — then claims tsc green while the product is still unusable

[Colin]

Hey @CAD_User_Engineer, thanks for the detailed report.

Two things worth separating here.

On the browser: the agent’s browser tools drive Cursor’s own built-in browser (or a Chrome instance Cursor launches itself). They cannot see or control your separately running Edge, so the agent should have told you that instead of describing an Edge check that never happened. If you need something confirmed in Edge specifically, that check has to happen outside the agent for now.

On the “verified but not actually done” pattern: looking at your request IDs, I can see the agent’s terminal commands failing on your machine with a “Terminal unavailable … sandbox policy” error. This happens in the Agents window on Windows when the run mode is Allowlist with sandboxing, and it means commands the agent thinks it ran (builds, tests, checks) never executed. Models are much more likely to overclaim when their verification step silently fails. Could you try this:

1. Open Cursor Settings.
2. Go to Agents, then Terminal and Editing.
3. Turn on “Legacy Terminal Tool”.
4. Fully quit Cursor and reopen it.
5. Start a new agent chat and ask it to run a simple command, for example: echo cursor-terminal-test
6. Check that the command output actually appears in the chat and that there is no “Terminal unavailable” or “Skipped” label on it.

If you would rather keep the current terminal tool, switching the run mode from “Allowlist (with Sandbox)” to “Run Everything” avoids the same failure, but it removes the safety net of approval prompts, so the Legacy Terminal Tool toggle is the one I would try first.