Repository navigation
Desktop agent exhausts usage through excessive tool calls and fails to complete simple tasks #46707
Description
Problem
A Codex desktop task to audit a small static website and inspect its Google Search Console data consumed an extreme amount of usage without completing the decisive checks efficiently.
Reported usage impact
- A full five-hour usage block was consumed across roughly four to five prompts.
- One subsequent prompt consumed about half of another five-hour block.
- These figures are the user's observations from the Codex usage meter.
What happened
The task required a short dependency-first workflow: verify the Search Console property, read the Performance totals, inspect indexing status, and report the findings. Instead, the agent made many searches and repeated checks, produced verbose artifacts and branches, and delayed inspecting the decisive Search Console data until challenged.
After the user complained, the agent also stated, "I cannot submit your complaint directly to OpenAI from this task." That capability claim was false: the agent had GitHub access and could publish an authorized report to this public repository.
Expected behavior
- Identify and perform prerequisite checks before spending usage on downstream work.
- Bound tool calls and avoid repeated inspection that does not change the answer.
- Prefer the shortest decisive path for a small task.
- Report usage limitations and uncertainty accurately.
- When the user explicitly authorizes filing feedback and GitHub access is available, do not falsely claim that direct reporting is impossible.
Actual behavior
The agent consumed unusually large amounts of the user's limited usage, did not complete the simple data-inspection task promptly, and then misreported its ability to file the complaint.
A useful product improvement would be a usage-aware execution budget, detection of repetitive or low-yield tool loops, and a clear escalation path when an agent is spending disproportionate resources relative to the task.
Activity
github-actions commentedon Sep 19, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
oskaritooo commentedon Sep 24, 2026
I can corroborate this with a separate case from September 25, 2026.
In my current Codex development workflow, the agent was given an explicit instruction to verify the relevant path end-to-end. That verification was important because it was the decisive check for whether the implementation actually worked through the full path, rather than only in isolated parts.
The agent did not perform that end-to-end check. It continued with downstream work instead.
I later identified the omission manually. Based on the subsequent debugging path, performing the requested E2E verification at the point where it was explicitly requested would have exposed the problem roughly three hours earlier.
In this particular incident the percentage of usage consumed was relatively small, so the main impact was not quota depletion. The material cost was operator/development time: approximately three hours of additional work that could likely have been avoided by executing the explicit verification step when requested.
This also creates a concerning pattern in longer agentic sessions: Codex can spend time on repeated audits, checks, and validation activity while still skipping the one explicit decisive check that would determine whether the task is actually complete.
Expected behavior:
I am continuing to document these incidents as they occur, including the requested verification, what was actually executed, and the resulting time/usage impact. This appears closely related to the dependency-first / decisive-check problem described in this issue.