QCon SF 2026: Why the Agent Harness, Not the Model, Decides What AI Agents Can Do
This summary and analysis were generated by AI from the original article at InfoQ AI and may contain errors (how Viqus works). Read the source for full details.
4
What is the Viqus Verdict?
We evaluate each news story based on its real impact versus its media hype to offer a clear and objective perspective.
AI Analysis:
The talk gives a useful vocabulary for agent engineering, but the core ideas about harnesses, permissions and feedback are already widely discussed, and its evidence is mostly borrowed from existing papers and model cards.
Article Summary
Manju Rajashekhar, founder and CEO of Mad Labs, opened QCon San Francisco 2026 with a keynote arguing that the harness surrounding an AI model, rather than the model itself, decides what an agent can do and how safely it does so. She defined the harness as everything in an agent that isn't the model, covering tasks, context, tools, state, permissions, feedback loops and a definition of done, and mapped its parts onto operating system functions: context as virtual memory, interfaces as system calls, execution boundaries as processes and permissions, and feedback as return codes. She cited benchmark evidence, including DeepSeek's report that one model scored between 65.5% and 74.2% on DeepSWE depending on the harness, and a Qwen3.8-27B result rising from 66.8% to 92.4% on hard LiveCodeBench problems inside a manager-worker scaffold. The talk extended her QCon AI Boston argument that production agents are distributed systems and should use an actor-model runtime. She closed by noting that harness components encode assumptions about model limitations that expire as models improve.
Key Points
- The harness, defined as everything in an agent other than the model, governs tools, context, permissions and feedback, and the keynote argues it matters more than the model choice.
- Benchmark examples cited in the talk show large swings from harness design alone, such as one model moving from 65.5% to 74.2% on DeepSWE across harnesses.
- Harness components encode assumptions about model weaknesses, so they must be revisited as models improve.
Why It Matters
This is a conference talk from a single speaker, summarizing engineering practices that are already circulating under the term harness engineering. The useful signal is the emphasis on measuring harness changes against model swaps, and on treating permissions as enforced boundaries rather than human approval prompts. Its industry-wide impact is limited, since it synthesizes existing ideas and cites prior results rather than announcing a new system or standard.

