Same Model, Different Harness: CUGA, Hermes, and Deep Agents — When to Pick CUGA?
Blog post from Hugging Face
IBM Research’s CUGA article argues that agent harness design can substantially affect an AI model’s performance and governance, reporting that the same GLM-5.3-flash model completed 81.1% of 90 AppWorld tasks with CUGA, compared with 46.7% using Hermes and 30.0% using LangChain Deep Agents. The comparison used a multi-app API benchmark and acknowledges that the harness configurations differed, the sample was limited, and no policy enforcement was measured in the benchmark. CUGA attributes its results to task decomposition, focused tool access, code-based actions in a sandbox, pre-execution verification, automated reflection, and pagination checks. The platform is aimed at enterprise automation, emphasizing runtime-enforced policies such as intent blocking, human approvals, tool guards, and controlled output formatting rather than relying solely on system prompts. CUGA is open source, supports hosted or open-weight models, can be deployed in private, sovereign, or air-gapped environments, and provides budgeting, run receipts, and configuration tools. The authors recommend alternative tools for terminal coding agents, personal always-on assistants, or general graph-based agent development, while positioning CUGA for governed business workflows spanning APIs, applications, and web interfaces.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| MCP | 2 | 8,906 | 856 | 195 | -25% |
| Kubernetes | 1 | 3,885 | 364 | 106 | -25% |
| OpenClaw | 1 | 513 | 30 | 17 | +63% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.