Devin: Cloud Tasks, Fusion, and a Review Workflow I Can Follow

On September 7, 2026, I used Devin on the Max plan through both the desktop and web apps. I sent work to a cloud agent and merged a pull request. What stood out was how closely task delivery, visible evidence, and review fit into the same workflow.

Devin's cloud workflow reminds me of Cursor Cloud Agents. In both cases, the useful handoff goes beyond a code diff: there is an opportunity to inspect what the agent did and how it checked the result. This is a similarity in how I work with them, not a controlled comparison of their success rates.

A handoff I can inspect

My impression is that Devin is eager to finish the assigned work properly and show the result through screenshots or video evidence. That makes the delivery easier to assess than a completion message on its own.

Devin's documentation describes a testing workflow that starts the application, follows a focused test plan, records the interaction, and returns a video attachment. Testing can be requested directly or started through the Test the app action. Cursor also documents screenshots, videos, and logs as cloud-agent artifacts. Those capabilities explain the resemblance I notice, without implying that either product produces complete test coverage on every task.

A recording is useful evidence of the interaction it actually shows. It does not establish that unrelated flows, edge cases, or the entire codebase are correct. Devin's own guide describes these recordings as focused checks and recommends existing test suites and CI for broader coverage.

Fusion reduces the model-selection work

Fusion is another part of the experience I appreciate. Instead of deciding which model should handle each piece of a task, I can leave that routing to Devin. My impression is that the output remains strong while making more economical use of models; I have not measured the savings on my own tasks.

Cognition describes Fusion as a combination of a frontier agent, a lower-cost sidekick, and model changes during a session. Its FrontierCode 1.1 Extended results, updated August 7, 2026, report frontier-level performance at up to 60% lower cost. That is Cognition's benchmark claim, not a discount on my Max subscription or a measured saving from this session.

The same report includes an example where delegation lost important task intent and reduced quality. For me, automatic routing removes a decision burden; it does not remove the need to judge the result.

Review before the merge decision

In the desktop and web apps, I found it convenient to follow draft and open pull requests from the Review area and initiate additional cloud reviews before merging. Keeping those steps close together made the process easier to manage.

Devin's Review documentation supports the underlying workflow: it lists open PRs, provides codebase-aware discussion, allows reviews to be triggered manually, and exposes draft, ready-for-review, and merge actions. Automatic review skips draft PRs until they are marked ready. Seeing a draft in the interface should not be confused with an automatic review having run on it.

I did merge a PR during this use, but I am not presenting the merge as proof that every review finding was correct or that every requested behavior was tested. The convenience is being able to inspect the proposed work and request another review before making that decision.

ACP brings other agents into the same interface

I also value Devin Desktop as an interface for terminal-oriented agents such as Grok Build and Amp. Working through ACP lets me keep these workflows in Devin Desktop instead of managing them only through separate terminal interfaces.

Cognition documents support for ACP-compatible agents inside Devin Desktop, and Grok Build's official documentation lists ACP as a supported way to use its coding agent in other apps. The ACP registry lists an Amp wrapper, so the Amp route includes an adapter dependency rather than assuming native ACP support.

This is a preference for a shared interface. Amp also offers native macOS and iOS apps, currently described as beta, so lack of an Amp GUI is not the reason to choose Devin Desktop. An exact Grok Build/Amp adapter configuration was not separately validated for this note, and I am not claiming compatibility across all versions.

What I take from this use

The appeal for me is the continuity: assign cloud work, inspect the evidence, review the PR, and decide whether to merge, with less model selection and interface switching along the way. This note records my September 7 experience on Max. It does not establish a general quality ranking against Cursor, a cost benchmark, or a reproducible ACP compatibility test.