Agentic AI: The Harness, Key to Execution and Risks

In AI agents, the model alone is not enough: execution, memory, safeguards, and observability are now treated as a discipline in their own right. Publishers are delivering ready-to-use frameworks while players like Databricks and Nvidia emphasize their crucial role in production, particularly for long tasks.

Production incidents primarily highlight the harness

Production feedback attributes the majority of failures to the harness rather than the model. Among the recurring issues are context degradation when exchanges lengthen and the absence of safeguards for irreversible actions. In regulated environments, compliance requirements make observability essential, through logs and traces explaining the decisions made. Research published by Nvidia at the end of August reaches similar conclusions for long tasks.

Choosing a harness becomes a distinct decision from model selection

The harness is now a separate choice from the model. Databricks reports that a solid harness around a mid-range model can outperform a weak harness associated with a more powerful model. The harness layer is thus placed on the same level as the model in the shift towards agentic systems. At VivaTech, Thibault Sottiaux (OpenAI) emphasized that a good agent results from the combination of a very good model and a good harness. Databricks anticipates changes but believes that execution environments, tool orchestration, and safeguards will remain crucial for maintaining production.

Structured offerings and practices, from Microsoft to Thoughtworks

Some publishers are now marketing pre-assembled harnesses. Microsoft includes this type of harness in its Agent Framework, which by default integrates task tracking, context compression, tool call validation, and observability, while allowing the disabling of features deemed unnecessary. Regarding usage, Birgitta Böckeler (Thoughtworks) differentiates between an internal harness, which accompanies an SDK or development tool, and an external harness, assembled by the user from instruction files, MCP servers, or specific skills. She also distinguishes between guides, which orient the agent before action, and sensors, which observe results and enable self-correction.

What constitutes an operational harness

A production harness relies on several complementary components. The system prompt continuously sets the role, objective, and rules of the agent. The model selects the tools, the harness executes them, and returns the results, often connecting through the MCP. Code execution occurs in an isolated, monitorable, and resettable sandbox. Memory and context management include context compression through summarization. Verification loops allow for testing, inspection, and review by the model. Safeguards impose permissions, policies, and human validation for irreversible actions, such as deleting a file or making a purchase.

Operation and scope: from ReAct to confusion with evaluation

The agentic cycle relies on a loop where the model reads the context, chooses an action, and the harness executes it until the task is complete. This framework was formalized by ReAct, while Toolformer demonstrated the calling of external tools; the vocabulary of harness came later. The term originates from software testing, which creates confusion with the evaluation harness, dedicated to measuring performance afterward, while the agent harness operates during the task. In terms of professions, harness engineering has emerged, encompassing prompt engineering and context engineering; the vocabulary established itself in early 2026, with disputed authorship. As models plan and correct their work better, some functions currently in the harness are expected to move back into the model.