The seven components
The harness uses seven components. They are not seven separate products; they are lenses for checking whether the workflow is complete.
- System Prompt — Active rules, boundaries, and the instruction to operate with PREVC.
- Tools — Tools that read, edit, validate, and request approval.
- Context Management — Loads only the relevant instructions, specs, and state.
- Verification — Commands and evidence that verify the result.
- Memory — Goals, handoffs, and durable artifacts, with retention.
- Sandboxes — There is no OS sandbox. Policy does not replace isolation.
- Hooks — Allowed integrations; they do not create autonomous execution.
What each component protects
| Component | Question it answers | Current limit |
|---|---|---|
| System prompt | Which rules apply in this session? | Rules do not replace human permission |
| Tools | Which action can be executed? | Mutating tools still require policy and approval |
| Context | What do I need to read before deciding? | Context must be relevant, not a global dump |
| Verification | What fact proves the result? | Evidence declared by the model is not a trustworthy receipt |
| Memory | How does the next session continue? | Retention and auditing do not prove completion |
| Sandboxes | Is the process isolated from the OS? | No. There is no OS sandbox installed |
| Hooks | What integrates into the lifecycle? | Hooks cannot start autonomous loops |