Agentic AI
Harness Launches Production-Ready Agents With Built-in Safety Controls
Autonomous Worker Agents let enterprises replace static pipeline steps with reasoning AI while maintaining the audit trails and governance they already rely on.
The New Stack
Harness unveiled Autonomous Worker Agents this week, a new capability that allows organizations to swap out fixed procedural steps in their software delivery pipelines with AI agents capable of making decisions on tasks like deployment, testing, and security scanning. Unlike earlier Harness agents designed to advise developers, these agents can directly execute production actions while remaining subject to the same compliance and audit frameworks enterprises already have in place.
The move represents a shift in how Harness views agent deployment. As CEO Jyoti Bansal puts it, "building an agent is becoming easier and easier and easier, but the harness is where the hard work is"—a reference both to the technical challenge of running agents safely at scale and to the company's founding vision of providing a "safety harness" around developer work.
Simplicity Through Markdown#
Defining a Worker Agent requires no specialized tools: developers write instructions in plain English within a single Markdown file, a format Harness describes as an industry standard. Teams can also have Harness generate the file automatically. Once deployed, agents access the Harness Software Delivery Knowledge Graph—an internal map of the customer's services, pipelines, deployments, incidents, and security issues—to inform their reasoning.
Safety by Design#
The critical difference between agents used for code generation and those running in production pipelines, Bansal notes, is that downstream systems can catch bad coding agent output, but there is no failsafe when the agent is the downstream system. Harness addresses this through several controls built into the platform over recent years:
- Isolated execution: Worker Agents run on Harness delegates—components hosted inside the customer's own infrastructure—meaning agents operate where production systems actually reside rather than in Harness' cloud.
- Sandboxing and scoping: Each agent runs in a restricted container with limited file and network access, its own identity, and permissions tied to the policy engine that governs human-triggered deployments.
- Auditability: Harness logs what triggered each agent, the prompt it received, the number of reasoning steps it took, and the outcome—treating agents as just another auditable action in an already-instrumented pipeline.
- Cost governance: Token spending is tracked per agent and per pipeline with budget caps that flag approval requests before costs spiral.
