Spawn autonomous subagents to run independent simulations in parallel.
Parallel Execution lets the AI delegate independent simulation tasks to autonomous subagents that run concurrently. Instead of executing parameter variations, solver comparisons, or mesh studies one at a time, the agent launches multiple workers that each handle a self-contained simulation from setup through results.
When to use Parallel Execution
Parallel Execution is useful when your workflow involves multiple independent simulation runs:
Parameter sweeps -- Run the same case with different inlet velocities, turbulence intensities, or mesh densities simultaneously
Solver comparisons -- Compare simpleFoam vs pimpleFoam or different turbulence models on the same geometry
Mesh convergence studies -- Run coarse, medium, and fine meshes in parallel to assess grid independence
Geometry variants -- Test design variations (e.g., different angles of attack or pipe diameters) concurrently
Parallel Execution vs. Batch & Sweeps
Batch & Sweeps orchestrate structured parameter studies with tabular output. Parallel Execution provides finer-grained control -- each subagent is a fully autonomous worker with access to the complete simulation toolkit, capable of adaptive decision-making during its run.
How it works
AI decides to parallelize
When a task involves multiple independent simulation runs, the AI launches one subagent per run in a single response. Each subagent receives a task description, a dedicated work directory, and optionally a shared directory containing common files (e.g., a base mesh).
Subagents launch
Each subagent is an autonomous AI worker with the full simulation toolkit -- file operations, mesh generation, solver execution, error recovery, and knowledge search. Subagents run concurrently and independently.
Progress updates stream in
As each subagent works, you see progress updates in the chat showing which tools are being used and what step each worker is on.
Results collected
When all subagents complete, the parent agent receives a summary from each one -- including files created, commands run, key results, and any errors encountered. It then synthesizes the combined findings for you.
Subagent capabilities
Each subagent receives:
Its own work directory -- All files are created inside a dedicated subdirectory (e.g., run_Re1000/, mesh_fine/) to keep runs isolated. Strict file discipline prevents writes outside the assigned directory.
Full simulation toolkit -- The same tools available in a normal simulation conversation: command execution, file writing, progress tracking, and knowledge search. The subagent inspects its available tools and skill catalog to decide which capabilities the task requires.
Shared directory support -- When specified, the subagent copies files from a shared directory using a mandatory copy-first pattern (cp -r shared/. workDir/). The subagent inspects the copied files before loading additional skills.
Host tool catalog awareness -- Each subagent receives a catalog of the parent session's available tools, letting it reason about end-to-end capabilities while only calling tools from its own toolset.
Autonomous skill loading -- Subagents decide for themselves which skill files to load from the skill catalog. They follow a local reference order (skills, then agent resources, then knowledge search).
Solver consistency rules -- Each subagent follows the same OpenFOAM configuration rules as the main agent (e.g., blockMesh and checkMesh must succeed before running a solver)
Subagents run with a default limit of 30 tool steps, configurable up to a maximum of 500 steps.
Workspace isolation
Subagents follow strict file discipline:
All created files go inside the assigned work directory
No modifications outside the work directory boundary
When a shared base is provided, the copy-first pattern is mandatory -- the subagent copies the shared directory into its workspace before making any edits
After the initial copy, all commands run with cwd set to the work directory
Example
Ask:
"Run this backward-facing step case at Re = 1000, 5000, and 10000. Compare the reattachment lengths."
The AI will:
Set up a shared base case with the common geometry and mesh
Delegate three subagents (run_Re1000, run_Re5000, run_Re10000), each copying the base case and adjusting the inlet velocity
Collect reattachment length results from all three runs
Present a comparison table
Availability
Parallel Execution is available on all plans. Each subagent consumes tokens from your plan's token budget, so parallel runs use tokens faster than sequential runs.