Episode 93 · Harness Design series
The Loop and the Hands: What a Coding Harness Is (Harness Design, Part 1)
Platform engineer Ines Varga doubles her coding agent's success rate without changing the model, and her mentor Gideon Achterberg uses it to explain the harness: the loop, the tools and the guard rails around a model with no memory. We trace the history from ReAct to SWE-agent's agent-computer interface, tour OpenHands, Aider, mini-swe-agent and the commercial agents, then go inside the harness from 'An Empirical Study of Harness Design for Coding Agents' (arXiv:2609.20804). In that study, predefined tools scaffold weaker models (worth 15 points on SWE-Bench Verified for Nemotron-3 30B), while a bash-only interface made Nemotron-3 550B slightly more accurate at half the cost, and Mistral Medium 3.5 needed tools for repo editing but preferred the shell on Terminal-Bench. The episode also covers granularity, tool design, how to read the statistics, safety gates, stuck detection and sandboxes.
Listen to the show on your preferred podcast platform.
Inside this episode
- Cold open: same brain, different hands
- First principles: a model is a function, a harness makes it an agent
- A short history: from reasoning traces to agent-computer interfaces
- The agent-computer interface, and why models need their own
- The harness family tree, in twenty twenty-six
- Inside the paper's harness, piece by piece
- The tools on the bench
- The action-space result: scaffolding for some, overhead for others
- Granularity: how the hands shape the code
- One ticket, two pairs of hands
- Designing a tool a model can actually use
- Reading the action-space numbers with care
- Gates and guardrails: what the paper held still
- Sandboxes: where the hands are allowed to reach
- What this means, and where we go next
Production credits & AI disclosure
Scripts generated with Claude Opus. Synthetic narration generated with Edge TTS. Reviewed, listened to, and approved by Sankalp Thakur.