DoorDash's reviewer approved 35,000 changes without reading them

On October 23, DoorDash ran 35,000 code changes through a multi-agent system at $4.79 each. The reviewer who approved or rejected each change read a summary the agents had written about their own work. The reviewer did not inspect the code. The handoff document was the product, and its accuracy was the thing that either earned or lost the reviewer's trust, at scale, with no secondary check on whether the summary matched what the agents had actually done.

The same day, a paper on arXiv showed that a group of agents can reach high collective confidence by copying the first agent's answer rather than by independently verifying it. A reviewer looking at that group's output sees agreement; the review surface shows no record of where the agreement came from. The DoorDash reviewer and the arXiv reviewer share the same exposure: the document in front of them looks authoritative, and nothing in it marks the seams where that authority was constructed.

Atlas, a tool for replaying long agent runs, was announced the same day to address a different version of the same gap. The agent had finished; the human arrived after. Atlas shows what the agent did, in sequence, so the human can judge the run rather than just its output. The design insight Atlas names is narrow: the review surface is a separate artifact from the work itself, and building it after the fact is harder than building it during the run. Most products ship the output and leave the review surface as an afterthought, which means the human who arrives after the agent finishes has the summary the agent chose to surface, and nothing else.

A third paper, published the same day, found that a writing scaffold placed before generation improved the quality of AI-assisted writing in a randomized controlled trial. The scaffold changed what the writer knew before the AI started, which changed what the writer could judge after. The reviewer's position at the end of the run depended on what they understood at the beginning. That sequence matters: a human who delegated without framing the work cannot review the output with any footing, because they have no prior position to compare it against.

Taken together, these four October 23 findings describe a single failure mode. The reviewer is handed a document. The document was written by the thing being reviewed. The reviewer has no view into how the document was composed, which agents contributed, whether agreement among agents reflects independent reasoning or simple propagation, or whether the summary omits anything the agents found inconvenient to surface. Build the review surface before you build the delegation flow: decide what artifact the reviewer will receive, what that artifact must show about its own provenance, and how the reviewer will mark their judgment on the record. If your team cannot answer those three questions, the approval your product asks for is not an approval. It is a signature on a document the reviewer was not positioned to read.