Models, Agents and harnesses
Separate generation, delegated responsibility and the runtime that makes it durable.
A model transforms input context into generated output. It does not independently preserve a goal, operate tools or own the consequences of completion.
A chatbot adds conversational turns. A copilot adds instructions, retrieval and workflow integrations while the human still carries the goal. A tool-using Agent crosses the delegation threshold when it can own a bounded outcome through a reason–act–observe loop and return evidence.
A harness makes that Agent operable. It owns session lifecycle, context assembly, tool access, permissions, sandboxing, model selection and recovery. Pi and Codex are distinct harnesses; an AgentOS Assignment is not inherently a Pi session.
Example
“Suggest a validation check” can be answered by a copilot. “Ship onboarding validation through this project’s reviewed workflow” is a bounded Agent outcome. The harness may resume context, invoke Git and tests, and expose the session. AgentOS adds the durable owner, supervision and authority boundary around that Agent.
Only harnesses and launch shapes verified by the selected AgentOS release should be described as supported. Harness-native settings remain opaque inputs rather than a schema AgentOS pretends to own.
Last updated on