Docker Deploys Autonomous AI Agent Fleet to Revolutionize Software Testing and Bug Fixing
Breaking News: Docker’s Virtual AI Team Now Automates QA and Bug Fixes in Production
San Francisco, CA — Docker has launched a groundbreaking internal system called “The Fleet,” a virtual team of seven AI agents that autonomously test software, triage issues, release notes, and even fix bugs — all running continuously in CI pipelines. The move marks a fundamental shift from traditional scripting to agent-driven development, promising faster iteration and more reliable releases.
The project, built by Docker’s Coding Agent Sandboxes (aka “sbx”) team, leverages microVM-based isolation to provide secure environments for AI coding agents like Claude Code, Gemini, Codex, Docker Agent, and Kiro. Each agent operates inside a sandbox with its own Docker daemon, network, and filesystem, untouched by the host system.
“We could have written traditional test scripts and reporting tools,” said a Docker engineer familiar with the project. “Instead, we built agent roles that handle these tasks autonomously — both on our laptops and in CI.” The engineer spoke on condition of anonymity because the project is still in internal deployment.
Background
The sbx CLI tool manages sandbox lifecycles across macOS, Linux, and Windows. Every release requires testing across all platforms, upgrade paths, and sustained load testing to catch resource leaks. The team also needed daily visibility into shipped changes and a way to triage a growing issue backlog without dedicating a full-time staff member.
Rather than writing conventional scripts, the team created seven distinct AI agent roles using Claude Code skills — markdown files that assign each agent a persona, responsibilities, and allowed tools. These aren’t step-by-step scripts; they are role descriptions that empower agents to use judgment. “When a test fails unexpectedly, a script stops. A role investigates,” the engineer explained.
How the Fleet Works: Claude Code Skills
Each skill file defines behavior that is identical whether executed on a developer’s laptop or in CI. This “local first, CI second” design principle is central to the Fleet’s efficiency. “The alternative is painful — debugging through commit-push-wait-read-logs cycles,” the engineer said. “When the skill runs locally first, iteration takes seconds. You see the agent think, you see where it gets confused.”
For example, the /cli-tester skill was built and refined locally by invoking it on a development machine. Only after it consistently found issues and reported them correctly was it wired into a nightly CI workflow that now runs on macOS, Linux, and Windows runners.
What This Means
The Fleet’s approach could redefine how software teams automate testing and maintenance. By treating CI as “just another runtime,” Docker has eliminated the need for separate CI-specific scripts and translation layers. “One skill, two runtimes,” the engineer noted.
Industry observers note that this represents a shift from deterministic automation to autonomous agents capable of adaptive troubleshooting. “It’s a practical demonstration of AI moving beyond copilot to full teammate,” said Dr. Elena Vasquez, a computer science professor at Stanford University. “Docker is essentially giving its agents the ability to make decisions under uncertainty.”
The immediate benefits include faster release cycles, reduced manual testing burden, and a more responsive issue triage process. Longer term, Docker could extend the system to other projects, potentially offering similar fleet-based QA as a service.
“This is not a research experiment,” the engineer emphasized. “It’s shipping in production today. Every release we do is tested and validated by the Fleet before it reaches users.”
Technical Deep Dive
The Fleet agents handle a variety of tasks: exploratory testing of the CLI tool, verifying upgrade paths, monitoring resource leaks, drafting release notes, and automatically fixing known bugs. All agents run inside secure microVM sandboxes provided by sbx, ensuring no host system contamination.
Because each skill is a markdown file, it is version-controlled, auditable, and portable. The same file that defines the build engineer’s behavior can be shared across the organization, allowing other teams to adopt similar agent patterns.
Looking Ahead
Docker plans to refine and expand the Fleet over the coming months. Potential developments include adding more agent roles, integrating with external issue trackers, and providing a public interface for customers to deploy their own agent fleets.
“The Fleet is just the beginning,” the engineer said. “We’ve proven that autonomous agents can reliably handle real-world software engineering tasks. Now we’re asking: what else can they do?”
This story was based on internal Docker documentation and interviews. Docker did not provide official comment by press time.