Agentic AI
Enterprises can deploy unified AI code frameworks across teams, study finds—solving governance gap in agent adoption
Researchers propose a harness-based architecture that lets large organizations run identical AI-agent code across multiple deployments while maintaining central governance and auditability.
Cost has plummeted for generating custom code with frontier AI models—a specialized problem that once took weeks now takes an afternoon. But reviewing, maintaining, and governing that proliferating code remains expensive. Each bespoke solution drifts independently, forcing teams to re-read entire codebases to understand how they differ.
The governance problem#
Large enterprises typically respond with two approaches, neither ideal. They either standardize on a pre-built product with limited flexibility, or build a graph-based orchestration framework that feels custom-built for each use case but remains isolated in scope. Neither option fully escapes the tension between centralized control and operational flexibility.
Recent research shows a third path exists: the harness paradigm, which treats a coding agent's backbone as enterprise infrastructure rather than a customizable tool. Several studies have established that:
- Harnesses perform adequately at the task level and actually outperform more complex architectures on real enterprise work
- The harness itself—not the underlying model—accounts for most performance variance in enterprise benchmarks
- The missing piece between technical viability and widespread adoption is governance
The proposed solution#
A team led by Juraj George Salapa proposes an architecture that closes this gap. The key: one standardized harness runs unmodified across every deployment. Because the code remains identical, auditing what gets built reduces to reviewing a single instructions file.
The framework introduces four governance mechanisms:
- Credential-scoped tooling: Each backend gets a single generic request tool paired with a limited-scope credential, eliminating the need for handwritten integration methods
- Authorization outside the harness: Deployment logic sits outside the core agent, letting the same artifact function as a scheduled task, a chat interface, and a terminal tool
- Push-triggered registration: Adding code to the repository automatically registers new capabilities—audit trails become simple text-file reviews
The reference implementation builds on microcc, an open framework designed for this architectural pattern.
