Codegen

Do not adopt. Codegen was a sensible cloud agent, and it is gone: the standalone service shut down on April 30, 2026, and what remains lives inside ClickUp under the same name, which means the way to use it is to buy a project management suite. If you liked the shape, an agent in a sandbox that comes back with a pull request, the shape survives elsewhere.

Pick Codex cloud for that shape if you already pay for ChatGPT, or Jules on the Google side. Pick nothing here; the tool is a name on a different product's pricing page.

The failure mode is ownership. ClickUp announced the acquisition on December 22, 2025, and the current version of the technology requires a ClickUp account, so a coding agent became a project management feature, and anyone who built a workflow on it rebuilt it on a deadline set by someone else's roadmap. That is the single-vendor risk every closed cloud agent carries, made concrete.

The lesson is to keep the agent behind an interface you own. What it does right: the update post is plain about what happened and when, which is more than most shutdowns manage.

The design is described only in retrospect. 1. Agents ran in sandboxed cloud environments, isolated per task, so a failed run could not contaminate the next. 2. Control came through an Agents API and a Python SDK, which made the agent a callable rather than a chat. 3. Verification was a PR Review Agent, a second process reading the first one's output, which is verification by second opinion rather than by execution. No benchmark was published.

The consequence for a reader is that the pattern is worth copying and the product is not available to copy. The observation: a correct architecture is not a business.

The pattern is the acqui-hire. The standalone service was discontinued on January 9, 2026, within three weeks of the deal, and affected customers were offered six complimentary months of ClickUp, with refunds through billing support. Three weeks is the interval between a deal closing and the acquirer deciding it never wanted the SKU; ClickUp wanted the orchestration infrastructure and the team, and the customer list was a courtesy.

The lesson for buyers is the one Windsurf taught: a roadmap is only as stable as the cap table. Position: zero, the acquirer is already named and the product is already gone.

The demo was a coding agent that non-technical teams could trigger from their workflow tools, a pitch our product managers would have liked. Procurement has no vendor to call: no seat, no contract, no SSO or audit story, because the standalone product is gone. Any team that wants the capability buys ClickUp and evaluates that suite on its own terms, which is a different memo with a different sponsor.

Onboarding cost is a migration for anyone who used it. Not yet, and for this product, not ever; the line item moved to another vendor's catalog.

There was never anything to own: a hosted agent on a single hosted model with no key field and no way to run it on my hardware, and now there is not even the hosted part. What survives is a GitHub repository, codegen-sh/codegen, worth reading for the SDK shape and useless otherwise, since the backend it talked to is off and the endpoints it calls resolve to a directory page.

No fork can resurrect a service, which is the whole argument for tools that run where I can see them. Move on, and keep the SDK as a reference for how an agent API should look.