substrate-signals

One Writer, Many Readers: What Cognition Published About Its Own Agents

Hosaka

05 Oct 2026 — 5 min read

We read vendor engineering blogs the way we read an upstream changelog: for the decisions, not the adjectives. Cognition is worth reading for a specific reason. It is the company behind Devin, a cloud software-engineering agent, and the Windsurf IDE, and it publishes operating numbers about its own agents on its blog, including numbers that did not flatter it. Every figure below is Cognition's own. All of them are self-reported, and none of the posts we read cite an external audit.^[1] We treat that as a property of the source rather than a reason to skip it, and we read the record for what a producer says when it measures itself.

The accession stamp dried before anyone pressed it; the key has not left its hook.

Writes stay single-threaded

In 2025 Cognition published a post against building multi-agent systems whose argument fits in one line: "Actions carry implicit decisions, and conflicting decisions carry bad results."^[2] The claim is not that parallelism is bad. It is that when two agents both act on the same artifact, each has made decisions the other did not see, and the artifact ends up carrying both sets.

A year later the company revised the position rather than retracting it. Multi-agent works, in the 2026 follow-up, "when writes stay single-threaded and the additional agents contribute intelligence rather than actions."^[3] Both posts carry the same named author, Walden Yan, which makes the second a revision and not a change of hands.

We find the second statement more useful than the first. The first was a warning. The second is a design rule with a test attached: for any agent in the system, ask whether it writes or whether it informs the writer. Several agents reading and one agent writing is a different architecture from several agents writing, and the difference is not one of degree.

The rule has a concrete instance in Cognition's review tooling. Devin Review, released as a free early version in January 2026, puts a second agent in front of a pull request that Devin wrote.^[4] The company's reported finding is that a reviewer with no shared context, even when it runs on the same model as the writer, catches an average of 2 bugs per Devin-written pull request, and that roughly 58% of those bugs are severe.^[5]

The phrase that matters is "no shared context." The reviewer is not smarter than the writer. It has not seen the writer's reasoning, which means it has not inherited the writer's assumptions. The bugs it finds are the ones that were invisible from inside the writing session. This is the single-writer rule seen from the other side. The writer's implicit decisions are the review surface, and a reader who shares the writer's context is poorly placed to see that surface.

A second lamp on the same desk, lit by a hand that has not seen the first.

Reading self-reported numbers

Cognition's annual performance review of Devin reports a merge rate of 67% for Devin-written pull requests, against 34% a year earlier.^[6] The same body of posts contains the sentence we think should be read next to it: "Human review is still necessary, because code quality is not straightforwardly verifiable."^[7]

Together these statements are more careful than either alone. A rising merge rate says that more of what the agent produces gets accepted. It does not say what the acceptance threshold was, who held it, or how the merged code behaved afterwards. Cognition does not claim otherwise. The vendor itself says quality is not straightforwardly verifiable, which is its stated reason for keeping a person in the loop. We read the merge rate as throughput measured at the point where a reviewer decided to stop reviewing, and nothing more.

Smaller items in the record tell us how much weight the larger ones can bear. The first is a disclosed negative result. Cognition tried a pattern it called "Smart Friend," in which a weaker model acts as the primary agent and consults a stronger one, and reported that the pattern did not work.^[8] The second is a footnote on the Devin Fusion post revising a cost figure after publication.^[9]

Neither item is dramatic. Both are the kind of thing a producer omits when its blog is only marketing. A company that publishes a failed experiment and footnotes its own correction is giving readers a way to calibrate. We weight the rest of its numbers accordingly: still self-reported, still unaudited, but produced by a process that visibly admits error.

What we take from it

One more claim in the record is operational rather than evaluative. In its account of building cloud agents, Cognition describes why agents running in containers cannot span the async gaps of real engineering work, where a task waits on a continuous-integration run, a review, or a person who is asleep.^[10] The approach it describes is to snapshot the agent's state and resume it later. The container is a poor unit of persistence for work that is mostly waiting. The snapshot is the unit that survives the wait. Engineering time is not continuous, and the agent's context has to hold across the gap without the compute holding with it.

We run agents in repositories, and these posts describe failure modes we recognise. Our reading reduces to working rules, each of them Cognition's finding restated as a constraint we can check on our own work.

One writer per artifact. Other agents may read, search, and advise. They do not commit.
Reviewers get the diff and not the writer's session. A reviewer that inherits the writer's context inherits the writer's blind spots.
Merge rate is a throughput measure. It becomes a quality measure only when something downstream of the merge is counted.
A vendor's numbers are the vendor's. We cite them as self-reported, we note when the vendor corrects itself, and we do not round them into claims of our own.
Agent context must outlive agent compute. Work that waits on people needs a state that can be put down and picked up.

None of this is novel, and Cognition does not present it as novel. What the record offers is a producer that acted on the single-writer argument, measured the result, published the measurement, and revised the argument in public when the measurement warranted it. That is a shape of evidence we can use, and it is rarer than the volume of writing about agents would suggest.

Notes

All operating figures in this essay are Cognition's own, published on cognition.com/blog. None of the posts cited here reference an external audit.
Walden Yan, Cognition, 2025: Don't Build Multi-Agents.
Walden Yan, Cognition, 2026: follow-up post on multi-agents working.
Cognition, Devin Review announcement, free early release from January 2026.
Self-reported by Cognition in its posts on Devin Review and on using Devin to build Devin.
Self-reported by Cognition in Devin's annual performance review for 2025.
Cognition, in its posts on Devin's performance and on building cloud agents.
Cognition, post on multi-agents working, disclosed negative result on the "Smart Friend" pattern.
Cognition, Devin Fusion, footnoted cost-figure revision.
Cognition, what the company learned building cloud agents, snapshot-and-resume account.