HarnessHandbook
The Harness Handbook is a methodology for documenting an AI agent harness so that its behavior can be understood, audited, and safely modified [Harness Handbook]. A harness is the surrounding system that makes an AI agent operational: its prompts, tool wrappers, permission checks, state management, and execution logic. The methodology addresses the difficulty of reasoning about how such a system actually behaves.
A single behavior in an agent, such as confirming with the user before deleting a file, is not implemented in one place. It is spread across the prompt that instructs the model, the wrapper around the delete tool, the permission layer, the state that tracks the pending action, and the code that executes it [Harness Handbook]. To answer a question about how the agent behaves, or to change that behavior safely, a developer has to search the whole repository and risks missing one of the scattered pieces.
The methodology organizes documentation into three tiers that let a reader move from a high-level question down to the exact code, without an exhaustive search [Harness Handbook]. The first tier is a system overview showing how a request flows through the harness. The second decomposes individual behaviors into their responsibilities and dependencies. The third provides implementation traces linked to specific code evidence. The links are produced by static analysis that extracts facts from the code, reorganizes them by behavior, and generates the documentation while preserving the connection back to the source.
The stated benefit is that a reader, or a coding agent performing a modification, can start from a behavior question and arrive at verified code rather than guessing from a repository-wide search, which reduces the chance of missing a dependency during an edit [Harness Handbook].
The Harness Handbook sits in Techniques at Assess, positioned outer. It is an early research artifact led by Ruhan Wang at Indiana University with academic and industry collaborators, and its public repository has drawn little adoption so far [Harness Handbook repository]. The reusable idea is the tracing method, mapping observed agent behavior back to verifiable code through layered disclosure; the handbook itself is the source evidence for that idea. It is worth exploring for anyone building or auditing agent harnesses, but it is novel and unproven in practice, which places it at the outer edge of Assess.
