Chainguard puts coding agents in microVMs and moves CI into their loop
Workspaces and Checks enter beta with isolated execution, mediated credentials and signed test results that can travel with a commit.
An isolated machine for each agent
Chainguard has launched Workspaces and Checks in beta, pairing a remote environment for coding agents with a verification engine that agents can call while they work. The October 14 announcements target teams running agents in parallel on migrations, refactors and vulnerability remediation.
Workspaces gives each developer or agent a separate hardware-isolated microVM and synchronizes the local working tree, allowing developers to keep their editor and Git workflow. Chainguard says real credentials stay outside the workspace; tools needing credentials on disk receive encrypted, short-lived substitutes. Optional egress policies restrict outbound connections to destinations the team approves.
That is a concrete change in where an agent operates: execution moves away from the developer’s machine, with access boundaries enforced by the environment. It does not make every permitted action safe, and the network restriction depends on configuring the optional policy.
CI becomes a tool call
Checks brings test execution into the agent’s development loop. Paired with Workspaces, the agent session suspends while checks run, then resumes with structured failure logs so it can revise the change. Warm snapshots contain preinstalled dependencies and are rebuilt when the lockfile changes, according to Chainguard.
Each passing run produces a signed attestation tying together the environment, test result and commit. Chainguard says downstream merge gates can verify that evidence and accept the change without repeating the suite. This can make a test result portable across the handoff, but an attestation does not prove that the suite covers every bug or security requirement.
Beta availability, with claims still to evaluate
Both product pages label the services as beta Technology Previews. Chainguard’s promises of results in seconds and reduced compute spending are launch claims, rather than independently established gains across all repositories. Teams evaluating the products still need to assess their own checks, integrations and credential policies. The launch’s practical hook is the combination of constrained agent execution and verifiable feedback before a pull request reaches the merge queue.