SimPilot's AI agent translates chat intent into concrete tool actions. It combines model reasoning with a structured tool layer instead of relying on free-form text generation alone.
How the agent operates
At a high level, each request goes through:
- Memory retrieval — relevant cross-session memories are retrieved by semantic similarity and injected into context
- Intent classification — dual-path classifier (keyword + LLM) categorizes the request: Analytical/Q&A, New Simulation, Comparison, Reporting, or Knowledge/Theory
- Tool planning — choose relevant tools and order based on intent
- Execution loop — run tools, inspect outputs, adapt when failures occur
- Response synthesis — return results as message content and artifacts
- Memory extraction — after the session, reusable facts are extracted and stored for future sessions
The agent can invoke several tool families during a single conversation.
Workspace and documents
- Create and edit text, code, and spreadsheet documents
- Read back saved documents
- Render generated outputs as artifacts
File and case operations
- Read, write, and edit files across the simulation workspace
- Scaffold case directories and generate solver configuration files
- Configure meshes and generate execution scripts
CAD and geometry
- Launch an interactive browser-based FreeCAD session for creating and editing CAD geometry (STEP, IGES, STL, .FCStd). Files saved in FreeCAD auto-sync to the case directory.
- Generate geometry programmatically using CadQuery scripts (STEP/STL output) through the engineering pipeline
Engineering pipeline
- Plan and execute multi-phase engineering workflows (geometry, mesh, simulation, post-processing)
- Artifacts chain between phases: geometry → mesh → sim → postprocess
- See Engineering Pipeline for details
Simulation execution and recovery
- Submit simulation runs
- Parse runtime logs and detect failures
- Review and rewrite case files after failures
- Retry with updated configuration
Analysis workflows
- Design and execute structured engineering studies with experiment batching, evaluation criteria, and decision tracking
- Compare results across simulations and check for regressions
- Extract quantities from completed simulations
Case versioning
- Snapshot the current case state with full reproducibility metadata
- Compare two case revisions showing changes in mesh, boundary conditions, numerics, and solver settings with severity indicators
- Check if a revision complies with a validated baseline, returning a trust score and deviation list
Report generation
- Generate publication-quality PDF reports compiled with LaTeX. Uses a planner pipeline (planning, data gathering, figure generation, compilation) with domain context and optional web research for validation.
Dashboard visualization
- Generate interactive data dashboards with 20+ chart types (line, bar, scatter, heatmap, contour, surface3D, flowGL, and more). Dashboards support interactive controls (sliders, selects, toggles) for filtering and slicing data. Saved to the simulation job or study session for persistent access.
Parallel execution
- Launch autonomous subagents for independent simulation tasks
- Each subagent gets its own work directory and the full simulation toolkit
- Subagents run concurrently for parameter sweeps, solver comparisons, and mesh studies
Research and retrieval
- Look up API docs, configuration syntax, and code examples for third-party software
- Search the web for external docs, forums, and troubleshooting references
- Fetch content from a specific URL
- Run deep multi-source research with synthesis and citations
- Connect to external tools through MCP-based connectors
Planning and clarification
The agent always starts non-trivial work with a structured plan. That creates an objective, task list, and optional scratchpad notes so the workflow stays visible and traceable while it executes.
When required simulation parameters are ambiguous, the agent requests structured clarification before proceeding. This is used for missing inputs such as solver choice, mesh approach, boundary conditions, geometry source, fluid properties, time control, or desired outputs. After asking, the agent waits for the user's answers before continuing.
Model usage
The agent is model-agnostic and you can switch models at any time. SimPilot supports 15+ providers including OpenAI, Anthropic, Google, Groq, Mistral, and xAI with automatic cross-provider fallback — if your chosen model is temporarily unavailable, the system seamlessly switches to an alternative. Available models include GPT-5 family, Claude Opus 4.6, Claude Sonnet 4/4.5, Gemini 3 Pro, Grok 3, and more.
Memory
Cross-session memory can be toggled per user via the memoryEnabled setting. When enabled, the system extracts reusable facts after substantive sessions and stores them for future retrieval. Those memories are no longer preloaded into every prompt; instead, the agent queries them on demand with queryMemory when prior user-specific or team context would materially help. The tool supports scope: "personal" for user preferences, scope: "organizational" for team findings, or scope: "all" for both.
Personal memory categories are:
- software_preference -- Stable user defaults and tool/solver preferences
- parameter_insight -- Reusable methodological lessons that generalized beyond one run
- error_pattern -- Execution-backed recurring failures and the fix direction
- project_context -- Project-specific conventions, deliverables, or workspace rules
- general -- Other durable user habits that do not fit the categories above
Memories are scoped for precision:
- User scope -- Personal preferences and patterns that apply across all projects
- Project scope -- Facts specific to a project (promoted from workflow scope when confidence is high)
- Workflow scope -- Temporary facts tied to an active engineering workflow run
- Agent scope -- Role-tagged memories (setup, execution, analysis, validation) for structured multi-phase workflows
The system also extracts facts from file attachments (configs, logs, reports, images) when they contain reusable information. Memory retrieval uses vector-based semantic search with embedding similarity, falling back to recency-based retrieval when embeddings are unavailable.
Confidence annotations
The agent annotates physics decisions with confidence levels: high confidence (90-100%) proceeds automatically, moderate (70-89%) proceeds with explanation, lower confidence prompts you to confirm or choose alternatives.
Conversation context management
Two mechanisms keep the agent effective in long conversations:
- Rolling chat summary -- Every 10 messages (configurable), the system auto-generates a concise summary of the conversation so far and injects it into the system prompt. This preserves context about objectives, approaches tried, and current state without consuming the full message history.
- Tool output offloading -- Large tool outputs (over 6,000 characters, such as verbose simulation logs) are automatically truncated to preserve the first 2,000 and last 2,000 characters. This keeps the opening context (commands, headers) and closing results (metrics, exit codes) while shedding verbose middle sections.
Guardrails and access controls
SimPilot applies safety checks before running simulations, including plan-based access controls, runtime limits, command validation, and organization-level content policies.
Why this matters
This architecture gives you a conversational UX with deterministic execution primitives. You get faster iteration than manual setup while keeping a traceable workflow through tools, artifacts, and persisted run metadata.
State the physical objective and constraints clearly (geometry, boundary conditions, key metrics). The clearer your intent, the better the tool plan and generated case.