Chaos Engineering for AI Agents: The chaos-experiment Skill

Where this fits

A developer just pushed a new microservice to staging. They suspect it might crash under load or lose connections when a pod dies. Instead of leaving the IDE to open another dashboard, they ask their coding agent: "run a chaos experiment on the checkout service." The agent needs a structured way to discover infrastructure, pick a fault template, and launch the experiment — that's exactly what this skill provides.

Why agents benefit

• The skill gives the agent a step-by-step procedure (check infra → browse templates → list existing → create), so it never guesses at the Harness MCP API.
• It can enumerate available chaos infrastructures and templates, so the user doesn't need to know names or IDs in advance.
• It can reuse existing experiments instead of creating duplicates, keeping the chaos library clean and auditable.
• The version 2.0.0 pin and the harness-mcp-v2 requirement tell the agent exactly what to verify before running anything.
• It reduces back-and-forth: the agent chains discovery, creation, and validation in a single session.

Practical scenarios

Pod deletion drill before a release

Before shipping a new version, the agent runs a pod-deletion fault on the canary deployment to confirm the rollout controller recovers within the SLO window.

CPU stress on a latency-sensitive service

When a user reports degraded performance under peak load, the agent triggers a CPU stress experiment against the service and correlates the results with existing dashboards.

Network fault simulation for distributed systems

For a multi-region setup, the agent injects a network partition fault to verify that circuit breakers and retry logic behave as designed.

Add it to your agent workflow

Install with:

npx skills add harness/harness-skills --skill chaos-experiment

Then ask your agent: "Check my chaos infrastructure and list available templates for a pod-deletion experiment." The agent will call harness_list with resource_type: "chaos_infrastructure", browse templates, and finally create the experiment via harness_create.

Compare before adopting

Before adopting, compare against alternatives like LitmusChaos or manual kubectl chaos scripts. Check the skill's maintenance freshness (version 2.0.0, Apache-2.0 license) and whether the Harness MCP v2 server is already in your stack. Also confirm workflow fit — if you're not on Harness, this skill is less relevant to you.

Why it is worth tracking

With 104 stars and maintained by Harness itself, this skill has solid community momentum and a clear ownership signal. It's worth tracking if you already use Harness Chaos Engineering or plan to adopt it, and you can evaluate it against your resilience testing cadence. For more context, see the skill page at https://www.openagentskill.com/skills/harness-chaos-experiment.