Researchers from Google Cloud AI Research and Northwestern University published the Harness paper on Oct. 9, 2026, showing how changes to the software surrounding a language model can improve an AI agent without retraining the model. Gemini 3.5 Flash’s average score on the WorkBuddy benchmark rose from 73.4 to 80.8 while its weights remained frozen, demonstrating that the surrounding software can materially affect performance.

Google released the project’s code on GitHub under the Apache 2.0 license. The software wrapper examined by the researchers is commonly called a harness.

What an agent harness does

A harness comprises the code, configuration and execution logic surrounding a language model, including its system prompt, tools, sandbox, memory, rules and the loop through which it plans, acts and verifies results. Vivek Trivedi defined it in “The anatomy of an agent harness” as “all the code, configuration, and execution logic that is not the model itself.”

A language model alone accepts text and produces text. The harness gives it access to files, terminals, web search and browsers; manages its context and memory; launches sub-agents; selects models for particular steps; and runs deterministic checks such as tests, linters or requests for user approval.

A scientific paper by Sanderson Oliveira de Maciedo identifies four conditions for an agent harness: an iterative reasoning, action and observation loop; tools that can modify an external environment; task-based context management; and at least one constraint, check or deterministic action that operates without the model’s consent. Under that definition, an agent software development kit supplies components but does not become a harness until a developer assembles and runs the loop.

Examples include Claude Code, OpenAI’s Codex CLI, Google’s Gemini CLI, OpenCode and DeepSeek Harness. Anthropic renamed the Claude Code SDK as the Claude Agent SDK in September 2025 and described it as “the harness that Claude Code runs on.” DeepSeek similarly separates the model from the tools and plugins that let it operate through its harness.

Performance gains without a new model

LangChain reported in February 2026 that it kept the gpt-5.2-codex model unchanged while modifying its Deep Agents harness. Its TerminalBench 2.0 score increased from 52.8% to 66.5%, moving the agent from the top 30 to the top five in the ranking, according to the company’s account.

The changes included a planning, implementation, testing and verification loop; a module that forced a final verification pass; and a detector that flagged repeated edits to the same file. LangChain also varied reasoning budgets by stage. Running every stage at the maximum level produced a 53.9% score because of time-limit overruns, while the high setting delivered 63.6%.

Anthropic engineer Prithvi Rajasekaran described a three-agent harness in March 2026 that separated planning, code generation and evaluation. In a retro-game test, Claude Opus 4.5 without additional harnessing ran for 20 minutes and cost $9, but the main game mode did not work. The full harness ran for six hours and cost $200, producing a working application.

Harness engineering and its limits

Harness engineering seeks to convert repeated agent failures into rules, tools or automated checks. Mitchell Hashimoto, creator of Terraform, Vagrant and the Ghostty terminal, described the approach on Feb. 5, 2026, including adding failure-based rules to AGENTS.md and building scripts for recurring tasks.

OpenAI followed on Feb. 11 with an account of an internal beta product built over five months without engineers writing code by hand. The codebase contained about 1 million lines, while agents opened and merged about 1,500 pull requests. The team grew from three engineers to seven and averaged 3.5 pull requests per engineer per day.

The approach carries trade-offs. Anthropic’s test showed that a more elaborate harness can cost more than 20 times as much as a single run. Rules can also become obsolete as models change, while access to files, terminals and networks creates security risks. Permission modes, sandboxing and separating credentials from generated-code environments are therefore part of harness design; Anthropic says credentials are kept outside the environment where generated code runs.

Source: Incrypted