OpenAI Files Patent for Clickable Receipts Showing What Its Coding AI Did
Handing a coding job to an AI is easy; knowing what it actually did is the hard part. OpenAI's patent application tags every edit and command so a summary can link straight to the proof.

- Publication number
- US 2026/0367911 A1(US20260367911A1)
- Applicant
- OpenAI OpCo, LLC.
- Filed · Published
- Sep 11, 2025 · Nov 19, 2026
- Inventors
- Andrey Mishchenko, Katy Shi, Shao-Qian Mah, Hanson Wang, Tongzhou Wang, Alexander Embiricos
- Status when we published
- Waiting for an examiner
- Document
- 20 claims
- US classification
- 717/101
- Parent application
- Claims priority from a provisional application 63806537 (filed May 15, 2025)
What OpenAI's AI coding receipts actually do
Imagine you ask an AI assistant to fix a bug in your app, go make coffee, and come back to a note that says "Done, all fixed." Would you trust it? You have no idea which files it touched or what it ran along the way.
OpenAI's patent application goes after that gap. As the AI works, it tags each thing it does, like editing a file or running a test, with a small marker that points to exactly what happened. When the job finishes, you get a plain-language summary where each claim is clickable. Tap a line about a changed file and you see the exact edit. Tap a line about a test and you see the log of what ran, with the relevant line highlighted.
The goal is to let you check the AI's work in seconds instead of taking its word for it. A person still reviews the changes before anything goes live.
retrieving, after an agentic software development task is completed by an AI model, data generated by the AI model to complete the agentic software development task; identifying a plurality of citation tokens to content within the data …
Translation: After the AI finishes coding, it collects data and identifies specific markers pointing to what it did.
How each edit and command gets a clickable tag
When you give the system a coding job, it starts a sealed workspace (a container, basically a locked, disposable computer) holding a copy of your code. The AI plans small steps, edits files, and runs tools such as tests, all on that copy.
Along the way, the AI writes citation tokens: short pieces of tagged text that point back to one specific event. The filing describes two main types:
- File citations record which file changed and which lines were removed or added, in the style of a diff (a before-and-after comparison of code).
- Command citations point to a line in the terminal log (the running record of commands typed and the output that came back).
After the AI finishes, the back end pulls out all that data, finds the tokens, works out each one's type from its format, and sends your device a summary. The summary is split up around the tokens, so each one can become a button or hover card that opens the matching diff or highlighted log. The filing says tokens can also mark failures, with a reason.
The model is trained with reinforcement learning (trial-and-error training with rewards) to produce these tokens. The work happens on a branch of a code repository, and the last step is a pull request, which is a proposal for a human to review and merge.
… provide a summary of actions performed during of the agentic software development task to a client device based on the data, wherein the summary includes information pertaining to changes generated based on the plurality of citation tokens.
Translation: It sends the user a summary showing what changes were made using those clickable markers.
Why reviewing AI-written code could get faster
If AI assistants are going to take on whole coding jobs while you do something else, the slow part becomes checking their work. A summary you can only verify by reading every changed file saves little time. Links from each sentence to its evidence turn review into a quick spot check, which is what busy teams need before an AI touches real software.
The filing calls this an audit trail: a record of where each change came from, useful for catching mistakes or tracing odd behavior later. It describes a sandbox, a review step, and a pull request, but names no product. Even if you never write code, the idea applies to any AI that acts on its own: an answer that comes with receipts is easier to trust.
This is the eighth OpenAI filing we've tracked since July on agents that act for you, following a live AI progress timeline and calling outside tools unprompted.
Source: US 2026/0367911 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time OpenAI patent news breaks.
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings · 11 sheets
11 drawing sheets from US 2026/0367911 A1 · click any drawing to enlarge

No comments yet