How GaugeDesk works¶
You do not need these concepts to complete your first run. Use this section when you need to understand a permission, review, or deployment.
The basic workflow¶
- Build or select one fixed agent version.
- Add it to a project.
- Allow the files and tools needed for one run.
- Run the agent.
- Review the proposed changes.
- Keep, discard, or release the result.
Agent instructions and project work are separate¶
Use an edit chat to change an agent. Use a work chat to use the agent on project files. A project file or work prompt cannot silently rewrite the agent.
Permission is explicit¶
An agent does not receive access only because it knows a file name or has a tool installed. GaugeDesk checks the exact file, tool, action, purpose, and current approval.
If required permission is missing or invalid, GaugeDesk denies the action.
Review comes before acceptance¶
Agent output is proposed work. A person reviews it before keeping it. Sending content to another person or system can require a separate release review.
Fixed versions make work repeatable¶
A saved version cannot change. A project or website session uses one exact version until an authorized person selects another version.
Where work runs matters¶
Local, hosted, and independently verified runtimes have different trust limits. The model provider still receives prompts and allowed context in plaintext.