Harness

The 13 August 2026 launch page for «DeepSeek Harness» carries the equation that anchors the term: agent equals model plus harness, the model being what DeepSeek calls the soul of the agent, the harness being what lets it read its environment, call tools, and keep working under real conditions. Addy Osmani, Shubham Saboo and Sokratis Kartakis put numbers on the split in their May 2026 Google whitepaper: model ~10%, harness ~90%, where the 90% covers instructions, tools and «MCP», sandboxes, orchestration, hooks and observability. Their evidence is a benchmark result: on Terminal Bench 2.0, swapping only the harness moved an agent from outside the Top 30 into the Top 5. Their conclusion, "most agent failures are configuration failures", relocates agent quality away from model choice.

Bill Chen and Brian Fioca of OpenAI's Applied AI team had made the same decomposition in November 2025, splitting an agent into user interface, model, and harness, and arguing the harness is the underestimated part: it absorbs each model's training habits (the tendency to read everything before acting), latency, context compaction, tool integration, and sandboxing. OpenAI's answer was to position «Codex» as an integrated agent-plus-harness rather than a model, and to tell developers not to rebuild the layer each time.

DeepSeek's answer is the opposite shape: a harness defined by plugin composition and a single append-only log, given away under MIT the same week its inference pricing changed. Both positions treat the harness as the durable asset while models cycle.