The solver remains the solver
Natural language chooses and coordinates actions. EPA SWMM execution remains deterministic, CLI-runnable, and directly comparable with bare swmm5.
Vision · human control by design
From running a model to defending the result.
Agentic SWMM treats a simulation as more than an output file. It preserves the goal, environment, inputs, solver run, checks, provenance, memory, and human decisions needed to inspect the result later.
The initial goal vs. the real problem
When this project started, the goal was simple: make EPA SWMM easier to install, run, reproduce, and check through an agentic workflow.
But as development continued, it became clear that real hydrological modeling is not only about running a model. A modeling workflow should also remember what was done, what assumptions were made, what failed, what was corrected, and whether the results can be trusted.
That shift in thinking is the heart of Agentic SWMM. The runtime can coordinate the work, but model files, SWMM runs, QA checks, plots, calibration trials, provenance records, agent snapshots, and audit notes remain visible as reusable artifacts—never hidden inside a chat. The goal is not to replace SWMM or the modeller, but to make SWMM-based modelling easier to reproduce, audit, extend, remember, and trust.
“The agent may coordinate the workflow. The evidence must still stand on its own.”
What trust means here
Inspectable before intelligent.
The system earns trust through explicit boundaries, not through fluent answers.
The run carries its context
Version, git commit, environment, provider, model, permissions, prompt hash, tool-schema hash, skill hashes, and lifecycle are recorded with the session.
A preview is not called validation
Research previews, synthesis baselines, calibration support, and execution benchmarks each state what they establish and where the evidence stops.
The three-part solution
Execution, memory, and controlled refinement
Reproducible SWMM execution
Run SWMM in a deterministic, inspectable way. Canonical builder, runner, QA, plot, audit, upstream, and review stages emit traceable artifacts before outputs are treated as evidence.
Audit-based modeling memory
Keep immutable raw audit notes beside curated, Obsidian-compatible memory that summarizes assumptions, QA issues, missing evidence, failure patterns, and useful modelling lessons.
Controlled skill refinement
The memory layer uses these audited records to propose refinements to existing workflow skills — staying coupled to the current skill-driven framework.
Human review & safe AI
A key point is that this is not autonomous self-editing. The memory layer can surface recurring patterns and propose refinements to existing workflow skills, but accepted changes still require human review and benchmark verification before they take effect.
Because skills drive the workflow, every proposal stays coupled to the current Agentic SWMM framework—it cannot quietly rewrite itself. This keeps the modeller in control. It supports careful professional use without pretending that automation removes the need for engineering judgement.