Windsurf's Cascade agent is dead: Devin Desktop's hard cutover to Devin Local

Cascade, the local agent behind Windsurf, reached end-of-life on 1 July 2026 with no grace period, forcing teams onto Devin Local — part of Cognition's June rebrand of Windsurf into Devin Desktop, which now runs on the open Agent Client Protocol so Claude Agent, Codex, Gemini CLI and others can all sit inside one editor.

Cognition rebranded Windsurf as Devin Desktop on 2 June 2026 via a silent over-the-air update — plans, settings, extensions, and keybindings carried over untouched. What didn’t carry over: Cascade, Windsurf’s local agent, which Cognition scheduled for end-of-life exactly one month later, on 1 July 2026. There was no soft deprecation window. Any CI pipeline, automation script, or workflow rule that explicitly invoked Cascade needed repointing to Devin Local before the deadline or it broke.

What replaced it

Devin Local is a ground-up Rust rewrite with parallel subagents, and Cognition claims up to 30% better token efficiency than Cascade. More significant for teams running mixed toolchains: Devin Desktop ships supporting the open Agent Client Protocol (ACP, Apache 2.0 licensed), which means Codex, Claude Agent, Gemini CLI, OpenCode, and Junie can all run as first-class agents inside the same editor rather than being bolted on.

That’s consistent with what we flagged in an earlier signal on tool stacking: AI coding tools aren’t consolidating into one winner. They’re splitting into layers — an editor or terminal as the control surface, with multiple agents plugged in underneath depending on the task. ACP support is Cognition betting that being the best host matters more than being the only agent.

Why the hard cutover matters

A one-month deadline with zero grace period for a tool embedded in CI pipelines is an aggressive migration timeline, even by AI-tooling standards. Teams that treated Cascade as a stable dependency rather than a fast-moving product got a forced migration with no warning runway. That’s a recurring risk with AI dev tooling generally: the tools are improving fast, but “fast-moving” cuts both ways when your build pipeline depends on one of them.

So what

If any part of your delivery pipeline hard-codes a specific AI coding agent, that’s a dependency risk worth reviewing now, not after the next forced migration. We build AI-assisted workflows with that volatility in mind — see how on the AI-assisted development page, or talk to us about de-risking your own toolchain.

← All research

More research.

GitHub's Copilot metrics now count OpenCode and Claude Code traffic: who really uses which AI coding tool

On 23 October 2026 GitHub added a per-tool breakdown to the Copilot usage metrics API, so tools such as OpenCode and Claude Code are counted even without editor telemetry. It shows how mixed the AI coding tool stack inside software teams has become.

Read ›

GitHub Copilot retires Claude Opus 5 and GPT-5.6 Terra on 16 November: the model-churn bill for buyers

GitHub has announced that Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra leave every Copilot experience on 16 November 2026. For anyone commissioning AI-assisted development, model retirement is now a monthly event to plan around.

Read ›

GitHub Copilot now benchmarks how 'agent-first' your developers are against your industry

On 22 October 2026 GitHub added opt-in industry comparisons to its Copilot impact dashboard, ranking firms by share of agent-first developers and by the merge rate of agentic pull requests. Here is how to read the numbers when you are buying software.

Read ›

Apple's new bait-and-switch rule 2.5.19: what it means for AI-built and fast-shipped apps

On 16 October 2026 Apple updated its App Review Guidelines: apps that change after approval can be removed from the App Store and the Developer Program, and AI-generated intimate imagery is now explicitly covered. Here is what founders commissioning an iOS app should check.

Read ›

GitHub Copilot agentic autofix is generally available: AI now fixes security alerts, and bills for it

GitHub made agentic autofix for code scanning alerts generally available on 15 October 2026. Copilot can now fix up to 25 security alerts in one pull request, using AI credits your supplier or your team pays for.

Read ›

GitHub AI Scan for pull requests goes GA: AI security checks for PHP, Terraform and Dockerfiles

GitHub's AI Scan for pull requests became generally available on 15 October 2026, checking languages the standard scanner misses. Findings are advisory and cannot block a merge, which matters when you buy software.

Read ›

Start a project

Building with AI? Let's talk.

Tell us what you are building. We'll give you an honest view of the right approach.

Start a projectSee our work