Snyk put the Evo Remediation Agent into Open Preview with a change in who starts the work. Until now the agent ran in a developer’s CLI or ADE. Now an AppSec team defines a campaign (which vulnerabilities, which repositories, what schedule) and the agent works through them in a Snyk-hosted sandbox and opens pull requests. It checks itself with builds, available tests and rescans, and it “does not merge changes.” Its hardest number is a mergeable fix rate of 90.2% across 307 dependency-upgrade tasks, against 77.5% for Claude Code with Snyk findings and 59.6% for Claude Code alone, all on the same model.
That number is Snyk’s, from Snyk’s own VulnBump benchmark, and it needs two labels. The post names the model as the constant without naming the model. And the 90.2% belongs to the “CLI-based Remediation Agent,” the engine underneath, not to the campaign product being announced, which runs in a hosted sandbox on public package registries under a daily fix allowance. What survives independently is the definition. A fix counts only if dependencies resolve, the targeted vulnerabilities disappear on rescan, the application builds, and previously passing tests still pass. “A pull request that breaks the build, fails existing tests, or leaves the vulnerability in place adds review work and doesn’t shrink the backlog.” That is a better yardstick than fixes generated, and anyone selling a fix agent can be asked for it.
The catalog holds the ingredients and not the new thing. Snyk has 50 API pages. The Snyk Issues API carries the findings a campaign would target, the Snyk Orgs API matches the Organization the post scopes one to, the Snyk Pull Request Templates API sits on the output, and the Snyk MCP Server is the agent-facing door. No page in the record is named for campaigns. For an Open Preview gated through the account team, that is not a failure, but it does mean the feature exists today as a product rather than as a contract anyone can inspect.
The measurement is strong, with one hole in it. Snyk holds a Kin Score of 79.1, exemplar, with access clarity and operational transparency both at 92.1, while contract governance sits at 22.0. Agent Readiness is 43.7, agent-ready. Reversibility is documented and the MCP server is lit; idempotency and dry run are unlit. The agentic access profile counts 280 operations, 133 acting and 10 human-in-the-loop. Snyk’s own design keeps a person on the merge, and that is the right place for this agent. The record has less to say for the next agent, the one calling Snyk’s API with no such person behind it.