Harness

"Agent = Model + Harness" is the equation at the centre of a May 2026 Google whitepaper by Addy Osmani, Shubham Saboo and Sokratis Kartakis, with the model counting for roughly 10 percent and the harness for the other 90. Their evidence is a Terminal Bench 2.0 result where changing the harness alone moved an agent from outside the Top 30 into the Top 5. "Most agent failures are configuration failures."

The term names what surrounds the model: instructions, tools and MCP connectors, sandboxes, memory, orchestration, hooks, observability. Bill Chen and Brian Fioca of OpenAI's Applied AI team had already split an agent into user interface, model and harness in November 2025, and spent most of their talk on why the third part is hard: model-specific habits, context compaction, latency, sandboxing. OpenAI's answer was to sell Codex as an agent plus harness rather than as a model.

DeepSeek released its own harness as open source under MIT on 13 August 2026, the day DeepSeek-V4-Pro reached GA and three days before a new API price grid. Every capability in DeepSeek Harness is a plugin swappable by configuration, and everything the model sees lands in an append-only session log.

Sumeet Gayathri Moghe records a side effect: coding harnesses from the past year put interactive web applications within reach of non-developers. A month later Fred Cavazza reads the model/harness separation as the competitive axis itself, noting that Google's Gemini also mobilizes third-party models, Claude among them. His reserve stands: the reliability and profitability of Google's "collaborateurs synthétiques" at scale remain to be demonstrated.