An agent harness is everything that constrains and shapes an agent

An agent harness is everything that constrains, shapes, and defines what an agent can do. It has four components: tools (what the agent can call), environment (where those tools run), hard controls (architectural limits the agent cannot bypass), and soft controls (prompted nudges that steer but do not bind). Together, these four define the agent’s action space: the set of all things the agent can actually accomplish. Change any one component and the action space changes.

The distinction between hard and soft controls matters more than it first appears. Hard controls are built into the system: no network access means the agent cannot phone home; no keychain access means it cannot read your passwords. The agent literally cannot do the thing. Soft controls are instructions that tell the agent to prefer certain behaviors or follow certain procedures. The agent can still deviate. When security matters, rely on hard controls; when flexibility matters, use soft controls. A coding agent like OpenCode has broad tools but tight hard controls (sandboxed filesystem, no privilege escalation). A data science agent extends the same harness with a notebook kernel environment. A voice-first agent narrows the tools and changes the interaction modality entirely. See memory is part of the harness not a separate tool and agent actions and planning must be observable.