Give agents somewhere to work
AI agents are good at reasoning, but reasoning alone is not enough for operational work. Supply plans need to be recalculated. Mining scenarios need to run through an entire shift. Financial investigations need to preserve evidence and continue across multiple sessions.
PySmith gives every agent a persistent Python environment built for this kind of work.
The agent can install approved dependencies, load data, run analysis, call solvers, execute simulations, save checkpoints and resume later without rebuilding its environment from scratch. PySmith provides the execution controls around that workflow: isolated workspaces, resource limits, permissioned access, reproducible environments and auditable activity.
Instead of a stateless tool call, agents get a durable computational workspace.
Build agents that do not just respond, but continue the work.
Stateless agents, multi-day work
Most AI agents can generate a response or call a tool, but they are not designed to maintain a working environment across complex, multi-day tasks. Every new session may require reinstalling dependencies, reloading data, rebuilding intermediate state and reconstructing previous decisions.
Most AI agents are built for short conversations, but real operational work takes hours, days or weeks. PySmith gives agents a persistent, controlled Python environment where they can load data once, run analysis and simulations, use approved libraries and tools, save their state and resume when conditions change. With isolation, resource controls, policy-managed access and reproducible execution, PySmith helps teams build agents for supply planning, mining simulation and financial decision support without managing a full cloud environment themselves. PySmith is the persistent execution layer between an AI agent and real-world computational workflows.
Persistent. Controlled.
Persistent, policy-controlled Python execution for AI agents.
PySmith gives an agent a persistent, controlled Python workspace where it can execute real code, maintain state, use approved libraries, run simulations and resume work over time.
The workspace is still there
The agent can retain:
- Loaded datasets and validated dataframes
- Scenario configurations and simulation state
- Trained models and solver outputs
- Intermediate calculations and checkpoints
- Logs, decisions and previous observations
- Installed or cached Python dependencies
Execution inside explicit bounds
Work runs within operational boundaries:
- Per-tenant isolation
- CPU, memory, storage and runtime quotas
- Approved package and dependency policies
- Restricted network and secret access
- Sandboxed code execution
- Reproducible environments
- Audit logs and human approval checkpoints
Isolation is per-tenant system containers with resource quotas — not Firecracker. MCP is how agents find the desk; it is not unique to PySmith.
Load, work, checkpoint, resume
Agents route tool calls through MCP. PySmith is the always-on Python desk those calls land on — packages warm, state intact, work able to continue.
- 01Load data
- 02Analyse
- 03Run code
- 04Simulate alternatives
- 05Validate results
- 06Save a checkpoint
- 07Wait for new information
- 08Resume and replan
Load once. Keep the workspace. Resume when the facts change — not from a blank interpreter.
The connector is familiar. The difference is a durable workspace on the other side of python.exec, not a one-off tool call that evaporates.
Where the work actually lives
Evidence-led concepts, not claimed deployments. Each story shows where a persistent Python workspace sits — never the system that signs the decision.
Keep the plan when the facts move
An end-to-end supply plan is not a single forecast. When a supplier misses a date or demand shifts, planners need to diagnose the cause, test alternatives and replan without losing the assumptions already completed. PySmith keeps the planning computation available through every replan — people and solvers keep the decision.
Read the story →Run the shift, not the plant
An open-pit mine is a coupled system across a shift: trucks, shovels, crushers, breakdowns and production targets. PySmith hosts a persistent, inspectable Python environment for testing coordinated decisions across that horizon. It is decision support only. It does not control haul trucks, blasting, braking or other safety-critical equipment.
Read the story →Evidence that survives the session
Financial teams do not need a plausible paragraph. They need the applicable policy, the correct document version and a traceable explanation that can continue across sessions. PySmith keeps computation and review history together long enough to inspect, challenge and reproduce. It does not approve credit, execute trades or file reports.
Read the story →What PySmith is not
Not the forecasting model, optimisation solver, knowledge base or domain application. It is the execution layer that gives agents a reliable place to perform long-running computational work.
Not this: Run agent-generated Python in isolated, short-lived sessions. That is a per-second sandbox. Pause is the competing product.
Other products already sell burst scale, Firecracker sandboxes and GPU fleets. We will not print those as if they were shipping. The job here is a durable, controlled Python workspace for work that takes days.
PySmith is the persistent, controlled Python runtime for long-horizon AI agents.
From one-off tool calls to persistent computational workflows.
Build agents that do not just respond, but continue the work.
Request preview access. Docs are a stub while runtimes are in private preview. The shape will not surprise you: a connector, an instance, a workspace that persists.