A developer reported on 16 November that an approved MCP write through the OpenAI Responses API ran twice when the same response went on to a programmatic read. Teams that let AI agents change records in CRMs, content systems or finance tools through MCP servers risk duplicate edits, and OpenAI has not yet named a cause or a fix. OpenAI Developer Community bug report
Two writes in the server logs, one in the answer
The reporter, posting as tomwillmot, says server logs recorded two real tool executions in the order write, read, write. Both writes carried the same approval request ID and identical arguments but different item IDs, and the model’s final answer still claimed a single write. The generated program contained only a read. It happened after a successful write and after an MCP tool error. Control runs without the programmatic step produced one write each. The test used OpenAI’s GPT 6.1 Sol model, a synthetic MCP server and no real accounts.
OpenAI Support replied the same day that the team is “looking into this issue”. No root cause or timeline was given.
AI agent approvals do not stop duplicate tool calls
OpenAI’s MCP guide says each approval applies to one tool call. The sections we read give no advice on duplicate tool calls or idempotency, so one human approval tells you nothing about how many writes reached the server.
This has happened before. Microsoft’s agent framework issue 6851 described an approved tool running again on a later, unrelated turn when responses were chained with previous_response_id. A maintainer wrote that it was fixed in two pull requests. The new report uses stateless requests, so the trigger differs, and the reporter says a shared root cause is not established.
Our reading, which we have not tested, is that in both cases the platform treats an approval as state it can replay. The control you can rely on is a duplicate check at the write endpoint itself.
Limits and what is unconfirmed
The evidence is one reporter and a handful of failing runs, which the reporter says is not a rate estimate. The packaged test setup was checked only locally, and we have not reproduced it. Other models, streaming calls and production servers are untested, and OpenAI has not confirmed the bug.
Three stopgaps until OpenAI ships a fix
A community reply suggests three stopgaps, none endorsed by OpenAI. First, resume the approval with tool_choice set to none, which gave one write in the report’s control run, then send the read as a separate request. Second, require an idempotency key on every MCP write tool and have the server return the stored result for repeats. Third, flag any response that contains two calls with the same approval request ID. In the meantime, audit recent agent writes for duplicates and watch the thread for a fix.