01 / ISOLATION
Run safely
Give every workload a bounded execution environment with a clear lifecycle and explicit capabilities.
A small, deterministic runtime for running AI agents as isolated, observable, resumable workloads — without turning your platform into a distributed-systems science project.
Retries, isolation, timeouts, credentials, state, observability and failure recovery quickly become everyone’s problem. Dispatchable makes those concerns explicit primitives instead of framework folklore.
Give every workload a bounded execution environment with a clear lifecycle and explicit capabilities.
Treat execution as a durable state machine. A process dying should not mean your agent forgot what it was doing.
Structured events, checkpoints and policy boundaries make agent behavior inspectable by default.
Dispatchable keeps the core deliberately boring: accept work, establish a boundary, execute, emit state, checkpoint, and finish. Everything else composes around it.
Continue with GitHub →┌──────────────────────────┐ │ DISPATCHER │ └────────────┬─────────────┘ │ ▼ ┌──────────────────────────┐ │ WORKLOAD │ │ │ │ agent → tools → llm │ │ │ └────────────┬─────────────┘ │ ┌───────┴───────┐ ▼ ▼ CHECKPOINT EVENT │ │ └───────┬───────┘ ▼ RESUME / DONE
Designed for engineers who would rather understand the system than memorize another framework’s edge cases.
Security and resource constraints belong in the runtime contract, not scattered through application code.
Keep orchestration separate from the workload. Swap infrastructure without rewriting the agent.
Design around interruption, retries and partial progress instead of pretending they will not happen.
Read the model. Challenge the assumptions. Then ship.
Continue with GitHub →