Harness VegaLoad: Open-Source Load Testing Your Agent Writes

Harness argues that many services reach production with no load test at all, because the tests live in a separate tool, a separate repo and a ticket queue. Its answer, VegaLoad, is a free Apache 2.0 load tester built for a coding agent to write and run from the editor. For platform teams whose developers now ship agent-written code faster than performance testing can keep up, it is a bid to make a load test just another file in the pull request.

What Happened

In Harness’s launch post, Uma Mukkara, Head of Resilience Testing at Harness, introduces VegaLoad as an open-source project sponsored by Harness and built by its Resilience Testing team. The workflow is short: a developer asks their agent for a load test, the agent writes a test file next to the service code, runs it with VegaLoad, and returns a pass or fail result with one HTML report.

The tool ships as a single small binary for macOS, Linux, Windows, Docker and Kubernetes, with no JVM or Node runtime. It covers 14 protocols: HTTP/1.1, HTTP/2, gRPC, WebSocket, MQTT, Kafka, PostgreSQL, MySQL, Redis, RabbitMQ, FTP, SFTP, TCP and UDP. A built-in MCP server and agent skills work with Cursor and Claude Code, and vegaload init sets them up. The code is on GitHub.

Why It Matters

Two design choices make it relevant to platform teams rather than just individual developers.

Gates in the pipeline. Teams set limits for p95 latency, error rate and checks. A broken limit stops the pull request in CI or the rollout in CD, so a performance regression blocks a merge the way a failing unit test does.

Safety defaults for agents. An agent can test localhost freely, but any other host needs the developer’s approval, and every run is written to a local audit log. Tests run on the developer’s machine, and the report is a single HTML file that opens offline. In practice, that approval step means an agent can’t point a load test at a shared environment on its own.

What This Means for Platform Teams

The business model is familiar. Harness compares VegaLoad to LitmusChaos, the open-source chaos engineering project its team helps maintain: free tools that many teams will never outgrow, with Harness Resilience Testing as the paid path for distributed load from many machines, shared history across teams, and chaos resilience probes. Harness says it is building a path so VegaLoad tests move over without starting again. The tool is new, so maturity is unproven; the launch post makes no performance claims for VegaLoad itself.

Stack Archive’s coverage of CircleCI’s contract tests and agent evals in CI describes the same shift: checks that agents can run themselves, enforced in the pipeline.

Source: Harness Blog

Content created with AI assistance and reviewed for accuracy.