Workflows And Agents#

When you send a request, Agent Master does not answer from the language model alone. An orchestrator interprets the request, chooses the matching pyCSAMT workflow, collects any parameters it needs, runs the real agents, and returns the products with a provenance trace. This page describes that loop.

The Workflow Catalogue#

Asking what can you do? lists the workflows available on the loaded data. The core set is:

Workflow

What it does

qc

Quality control and per-station scan.

static_shift

Static-shift detection and AMA / LOESS correction.

denoise

RPCA / Hampel / AI denoising.

phase_analysis

Phase-tensor, strike, and dimensionality analysis.

tipper

Tipper / induction arrows.

rotation

Tensor rotation to a strike frame.

freq_decimation

Period selection / frequency decimation.

sensitivity

Sensitivity kernels / depth of investigation.

pre_inversion

Full pre-inversion preparation.

Beyond these, the orchestrator can run inversion (AI-neural, PINN, hybrid, and classic solver preparation such as Occam2D), report generation, and code generation (below). You do not call these by name — you describe the goal and the orchestrator routes to the right one.

Making A Request#

Describe the task in plain language, for example get the geoelectric strike. The orchestrator matches it to a workflow and starts running, streaming progress in the chat.

Agent Master executing the phase_analysis workflow with a progress bar

A request routed to phase_analysis, executing with a live progress bar (Executing phase_analysis… 4/5). Each workflow reports its step count and elapsed time as it runs.#

Choosing Lines#

When several survey lines are loaded, a workflow that acts per line first asks which lines to process — so a request like analyse only the line 22 is resolved explicitly rather than guessed.

The line-selection panel with the orchestrator result and figures behind it

The Select line(s) to process panel: toggle the lines you want and Run selected, or Run all lines. Behind it, a completed run reports Orchestrator routed … workflow ‘orchestrated_phase_analysis’ complete: 5/5 steps succeeded, with a steps trace and 5 figures generated.#

Providing Parameters#

When a workflow needs settings, Agent Master opens a small in-chat form instead of guessing — you stay in control of the science.

The Data Denoising parameter form with method, SNR threshold, and pipeline steps

A denoising request opens a Data Denoising form: pick the method (Wavelet / Median filter / Skip), set the SNR threshold, expand the Pipeline steps for period limits, then Run Workflow.#

These forms expose exactly the parameters the underlying agent accepts, with sensible defaults, so you can accept the defaults or tune them before the run starts.

Orchestration, Traces, And Cost#

Every completed request reports what happened:

  • the routing decision — for example Orchestrator routed to ‘report’ workflow (3 steps);

  • the outcome — Workflow ‘orchestrated_report’ complete: 3/3 steps succeeded in 3.3 s;

  • the cost of any LLM calls (often $0.000000 for local processing steps);

  • an expandable steps trace and a count of figures generated.

A completed report workflow followed by a sensitivity run, with the figures panel filled

A report workflow completes (3/3 steps, with a trace and two figures), then a analyse the sensitivity of the line request runs the sensitivity workflow — while the Figures panel fills with the thumbnails produced along the way.#

Chaining Requests#

Because the survey and the conversation persist, you can build up a session naturally — QC, then correction, then phase analysis, then a report — each request acting on the current state. You can also ask for a full pipeline in one message (run full pipeline: QC, inversion, and report) and let the orchestrator sequence the steps.

The agents Agent Master drives are the same ones documented in the agents user guide; the orchestrator that plans and sequences them is described in Workflow Orchestrator.

Next Steps#