5. Agent Master

Agent Master is the conversational application of pyCSAMT. You describe a CSAMT / AMT / MT task in plain language and it plans the workflow, asks for any parameters it needs, runs the right pyCSAMT agents, and returns figures, reports, reproducible code, and a full provenance trace. Its tagline says it plainly: describe your workflow — I handle the rest.

It complements the other surfaces rather than replacing them. The desktop GUI, web app, and MapView are point-and-click; Agent Master is the chat-driven front end over the same agents and the same scientific core, for when it is faster to say what you want than to click through to it.

A full Agent Master session at a glance: launch, configure the LLM, load EDI data, ask what can you do?, run a workflow (with line selection and generated figures), generate a validated script, and get a report with a sensitivity analysis. Each step below is documented in detail on the following pages.

A good first pass, in order, is Overview → Installation → LLM Configuration → Welcome & Chat → Workflows & Agents → Tools, Memory & Outputs → Agent Master & Other Apps → Troubleshooting.

• 5.1. Overview
  • 5.1.1. What Agent Master Does
  • 5.1.2. A Complete Request, Start To Finish
  • 5.1.3. When To Use It
  • 5.1.4. Run Agent Master
  • 5.1.5. The Interface At A Glance
• 5.2. Installation And Launch
  • 5.2.1. Install The App Extra
  • 5.2.2. Launch
  • 5.2.3. Options
  • 5.2.4. The Welcome Screen
  • 5.2.5. Verify
  • 5.2.6. Next Steps
• 5.3. Welcome Screen And Chat
  • 5.3.1. Welcome Screen
  • 5.3.2. The Main Window
  • 5.3.3. Loading EDI Data
  • 5.3.4. Chatting
  • 5.3.5. Sessions And History
  • 5.3.6. Next Steps
• 5.4. Workflows And Agents
  • 5.4.1. The Workflow Catalogue
  • 5.4.2. Making A Request
  • 5.4.3. Choosing Lines
  • 5.4.4. Providing Parameters
  • 5.4.5. Orchestration, Traces, And Cost
  • 5.4.6. Chaining Requests
  • 5.4.7. Next Steps
• 5.5. LLM Configuration
  • 5.5.1. The Settings Drawer
  • 5.5.2. Choosing A Provider And Model
  • 5.5.3. Privacy And Keys
  • 5.5.4. Running Without A Model
  • 5.5.5. Next Steps
• 5.6. Local language models
  • 5.6.1. Setup
  • 5.6.2. Limits and behavior
  • 5.6.3. Privacy and failure behavior
  • 5.6.4. Python use
• 5.7. Generating and editing scripts
  • 5.7.1. Follow-up edits
  • 5.7.2. Validation limits
  • 5.7.3. Python interface
• 5.8. Generated-code validation and repair
  • 5.8.1. Bounded repair
  • 5.8.2. Execution boundary
    • 5.8.2.1. Setting up Docker and the fixture image
• 5.9. Developer source questions
  • 5.9.1. Search boundaries
  • 5.9.2. Python interface
• 5.10. Tools, Memory, And Outputs
  • 5.10.1. Step Traces And Provenance
  • 5.10.2. The Figures Panel
  • 5.10.3. Generated Code
  • 5.10.4. Sessions And Pinned Prompts
  • 5.10.5. What To Keep
  • 5.10.6. Next Steps
• 5.11. Conversation context and stopping work
  • 5.11.1. Long conversations
  • 5.11.2. Answers and artifacts
  • 5.11.3. Stopping a request
• 5.12. Agent Master And The Other Apps
  • 5.12.1. One Set Of Agents, Several Front Ends
  • 5.12.2. Choosing A Surface
  • 5.12.3. Moving Between Them
  • 5.12.4. Failure Recovery
  • 5.12.5. Next Steps
• 5.13. Evaluating answers and scripts
  • 5.13.1. The benchmark
  • 5.13.2. Outcomes
  • 5.13.3. Running the benchmark
  • 5.13.4. Metrics
  • 5.13.5. Tests
  • 5.13.6. Limitations
• 5.14. Troubleshooting
  • 5.14.1. Launch
  • 5.14.2. Model Provider
  • 5.14.3. Data And Workflows
  • 5.14.4. Session And Browser
  • 5.14.5. When You Need More Detail
  • 5.14.6. See Also