5.7. Agent Master And The Other Apps#

The conversational, agent-driven experience is not unique to this app — it resurfaces inside the desktop GUI and the web app as well, because all three draw on the same pycsamt.agents package: the same Agent classes, and, for full-workflow requests, the same WorkflowOrchestratorAgent that plans a chain from a plain-language request. This page works out how the standalone Agent Master relates to those other agent surfaces, so you can pick the right one for a given moment of work and move between them without losing anything.

5.7.1. One Set Of Agents, Several Front Ends#

Nothing scientific changes between surfaces — routing, step execution, and the AgentResult returned by every step are identical wherever the request originates. What changes is framing: how much of the screen the conversation gets, and what sits next to it.

Surface

What it is

Agent Master (this app)

The dedicated, full-window conversational app — launched with pycsamt-agent. Best when the chat is the workflow.

Web app → AI Agents

A page inside the web app with an Agent Runner and a Chat tab, alongside all the point-and-click pages.

Desktop GUI

The desktop app opens an Agent Master panel window (Ctrl+Shift+A) alongside its Profile, Map, QC, and Inversion panels, so agent review sits next to the manual diagnostics it is commenting on.

Because the underlying agent coordinator and orchestrator are what actually run — see Workflow Orchestrator for how a plain-language request becomes an ordered agent chain — a survey processed through one surface behaves identically in another, and the products of a run (figures, reports, generated code) are interchangeable regardless of which window produced them.

5.7.2. Choosing A Surface#

The right choice usually follows from what is already open, not from any difference in capability:

  • Use Agent Master when you want a distraction-free conversation that plans and runs a whole workflow — and when you want generated, reproducible code or a report from a description.

  • Use the web app’s AI Agents page when you are already working in the browser studio and want to run an agent or ask a question without leaving it (see Processing Pages).

  • Use the desktop GUI’s Agent Master panel when your work is mostly manual review — QC plots, correction previews, inversion diagnostics — and you want agent commentary next to the evidence rather than in a separate app. As the desktop guide puts it, treat the agent as an explanation aid: let it point you back to the profile or QC view, then make the processing decision from the data itself, not from the agent’s confidence alone.

5.7.3. Moving Between Them#

Agent Master and the web app each start their own local server, and the desktop GUI runs as its own process again, so there is no shared in-memory session for a running agent to hand off — closing one does not hand a live conversation to another. What crosses freely is the same thing every surface already reads and writes: files.

  • Load the same EDI folder in whichever surface you move to. The EDI format and the loading behaviour are identical everywhere, so a survey that quality-checks cleanly in the web app opens the same way in Agent Master or the desktop GUI.

  • Carry generated scripts and exported figures across freely — a script produced by Agent Master’s code generation is a standalone pyCSAMT program with no dependency on the app that wrote it (see Tools, Memory, And Outputs), and an exported figure is just an image or HTML file.

  • Keep the survey folder alongside any saved session so the working context can be rebuilt on the next surface even though the conversation itself does not travel with you.

Because each surface binds its own port, you are not limited to one at a time either: Agent Master listens on 8765, the web app on 8050, and MapView on 8770 by default, so you can leave several running and move between browser tabs as the task shifts from chat-driven planning to point-and-click review.

5.7.4. Failure Recovery#

The same underlying stack means most handoff problems trace back to one of three places — the server, the model provider, or a mid-workflow parameter — regardless of which surface surfaced the failure.

  • If Agent Master is not reachable, confirm the pycsamt-agent process is still running and that its terminal shows http://127.0.0.1:8765 (see Installation And Launch); a closed terminal is the most common cause, since it is what hosts the server.

  • If a request needs a model and none is configured, add a provider and key in Settings and resend the request — you can still browse the interface without a key, but the language-driven planning and chat that make Agent Master useful will not run until one is set.

  • If a run fails partway through, read the steps trace for the failing step, adjust the parameters in the in-chat form, and run again (see Workflows And Agents). Because every step returns its own AgentResult, the trace tells you exactly which step stopped the chain rather than leaving you to re-run the whole workflow blind.

5.7.5. Next Steps#