Shepherd adds mandatory human checks to AI coding agents
Shepherd is an platform for managing such as Claude Code and Codex. The agent first turns a requirement into a design, which a person must review and approve before work can continue. It then divides the work into tasks and carries them out.
A second mandatory approval is required during before the work can be marked complete. The platform records measures such as how much work was completed by AI versus people and how often tasks succeeded. Agents pull work from Shepherd, allowing them to run on private internal machines without being exposed to the public internet.
The main goal is and a traceable work process, rather than making the agent itself more capable.
Key points
- A person must approve the proposed design before coding begins.
- A second human approval is required at the stage.
- The platform tracks AI-versus-human work and task success rates.
- Agents can run on private internal machines without public exposure.
- Direct token and cost tracking are not stated in the description.