The Anatomy of an Agent Harness

An agent harness is defined as everything in an AI agent that is not the Large Language Model itself. While the LLM provides reasoning and linguistic capabilities, the harness provides the "body"—the environment, tools, and context that allow the model to interact with the real world. Ryan Lopopolo, who coined the term, argues that developers should stop viewing agents as simple text-in/text-out interfaces and start treating them as autonomous systems that require deliberate environment curation.

Shifting Left: Automating Context and Quality

To achieve high autonomy, developers must "shift left" by automating the delivery of context and best practices. Instead of manually prompting the model with every requirement, engineers should build systems that allow the agent to discover documentation, verify code with static analysis, and access observability data (like PromQL) dynamically.

Key strategies for effective harness engineering include:

  • Context Efficiency: Structuring documentation with named anchors and links so agents can retrieve relevant information without overloading the context window.
  • Best Practice Enforcement: Using linters, tests, and coding conventions as automated guardrails that the agent must satisfy before a task is considered complete.
  • Iterative Loops: Moving from simple linear execution to closed-loop systems where the agent can observe its own failures, reflect, and retry, effectively mimicking the iterative nature of human software development.

Scaling Autonomy Through Diverse "Agent Teams"

One of the most powerful insights is the ability to curate a "diverse team" of agentic skills. By defining specific roles—such as a React architect or a performance expert—developers can compose agents that bring specialized expertise to a central task. This allows for the management of complex, ambiguous problems that would otherwise require multiple human specialists. As agents mature, they can be trusted with larger scopes of work, eventually enabling tasks like wholesale language migrations or long-horizon infrastructure management.

The Three-Layer Stack

Building high-performing agents requires a robust stack. The recommended approach involves:

  1. The Model Layer: Utilizing efficient, high-performance models like Gemini 3.8 Flash for high-frequency loops.
  2. The Orchestration Layer: Using tools like Antigravity Boost to manage agentic loops and state.
  3. The Knowledge Layer: Leveraging repositories like Google Skills to provide domain-specific context that ensures the agent understands the "golden thread" of what constitutes high-quality output for the organization.

Avoiding the Sunk Cost of Over-Scaffolding

Ryan warns against over-scaffolding—building overly complex, rigid harnesses that become obsolete as models improve. Instead, focus on two extensible interfaces: Tools and Context. By keeping these interfaces clean and standard, you can swap out underlying models as they evolve without needing to rebuild your entire agentic architecture.