Agent harness
The controlled agent runtime — policy, prompts, and tool discipline for steerable, auditable runs.
The agent harness is Studio’s controlled runtime for agent work. Models still generate; the harness decides how runs plan, use tools, compact context, and escalate when something needs a human.
What it governs
Admins configure the default profile under Settings → Agent Harness, including prompts, planning behaviour, tool discipline, and related model settings (for example compaction). Agents can point at a harness profile so specialist workflows stay inside organization rules.
Why it matters
- Steerable — long transformation work can be redirected without throwing away the run.
- Auditable — tool use and plan structure sit in a governed loop, not a free-form chat dump.
- Organisation-fit — prompts and policy match how your team decides, not a one-size consumer assistant.
Think of the harness as the trust layer around bring-your-own models: the same Stage and artefacts, with a runtime you can tune. It is not an eval lab — it is how reliable agent co-creation feels in the product.
Pair with Connect your AI, Agents, and Sessions for run history.