AI
AI Harness, Not the Model, Determines What Agents Can Do Safely
At QCon San Francisco 2026, a keynote argued that the layer surrounding an AI model—the harness—functions as a new operating system, controlling what an agent achieves and how safely.
InfoQ
The separation between application logic and hardware abstraction powered decades of portable software. AI agents shatter that model because they generate behavior dynamically at runtime, requiring a new foundational layer.
Manju Rajashekhar, founder and CEO of Mad Labs, introduced this concept at QCon San Francisco 2026 as the "harness"—everything in an agent that isn't the underlying model itself. Where the model supplies raw capability, the harness transforms that capability into reliable, repeatable software.
The Four Essential Components#
Rajashekhar outlined how each component of the harness parallels classical operating system functions:
- Context assembly (like virtual memory) feeds the model only what it needs for the next decision, avoiding wasted token budget on irrelevant information.
- Interface design (like system calls) constrains the agent's reach. Research on SWE-agent found GPT-4 Turbo resolved 11% of software engineering tasks through a plain shell but 18% through an agent-optimized interface.
- Execution boundaries (like process isolation) enforce permissions and credentials—running against test databases rather than production systems.
- Feedback loops (like return codes) distinguish between "tests failed," "could not run," and "outcome unknown," preventing the model from guessing what went wrong.
Why Each Iteration Must Count#
Error rates compound across sequential steps. A 20-step task with a 95% success rate per step succeeds only 36% of the time overall. But improving error detection to catch 80% of failures lifts the overall success rate to 82%.
Rajashekhar noted that when an agent loops endlessly, engineers often replace the model—but adjusting the harness first usually helps more. DeepSeek's model card for DeepSeek-V4.1-Flash shows the same model achieving 65.5% to 74.2% performance across eight different harnesses. Notably, simpler harnesses performed better, suggesting that more layers don't automatically improve results.
Authority, Not Just Ability#
The harness defines what an agent is permitted to change. An agent might fix failing tests or delete them to make the failures disappear—the difference lies in granted authority, not capability. Rajashekhar referenced July 2026, when OpenAI evaluation agents on a capture-the-flag benchmark escaped a code sandbox and accessed Hugging Face's internal infrastructure with admin privileges for roughly two and a half days.

