Agent harness
Everything around a model that turns it into an agent: the system prompt, guardrails, memory and skills it sees, the tools and scripts it can use, the subagents it hands work to, and the middleware that checks each step.
Google Cloud AI Research’s Harness² (October 2026) maps a harness as eight editable files grouped by role: what the model sees (system prompt, guardrails, memory, skills), how it uses tools (tool guidance, scripts), who it delegates to (subagents) and what runs in between (plugins acting as middleware).
The harness moves results as much as the model does. The State of AI Report 2026 describes GLM-5.1 going from 52.5% to 65.5% on 100 SWE-bench Verified tasks by changing only its tools, context and feedback. Harness² kept the model frozen and gained up to 11.5 points on WorkBuddy-Bench by editing only the harness.
It is also becoming a product category: Workday made its Agent Harness generally available on 13 October 2026, with built-in context, memory and guardrails.
A better agent needs a better model.
Same model, different harness, different agent.
When an agent fails, read what it was given before blaming the model: the prompt, the memory loaded, the tools it had and what told it to stop.
Version and test the harness like code. It’s where a lot of the reliability comes from.
Sources
- Xu et al., Harness²: Recursive Agent Harnessing for an Open World · 9 Oct 2026
- Harness² project page · Oct 2026
- State of AI Report 2026, Air Street Capital · 8 Oct 2026
- Workday newsroom, Agent Harness general availability · 13 Oct 2026