Long-running agent work has a lifecycle beyond a single conversation. A task may need human input, reach a context limit or be interrupted by a worker restart. Continuing that work requires a record of what happened, what remains and where execution should resume.
SEPT is an open-source Node.js/TypeScript durable task worker built around Pi’s native Durable Harness. Pi supplies the underlying agent runtime, including model access, authentication, generation and tools. SEPT manages the durable task lifecycle: persistent state and evidence, waiting and resuming, cancellation and restart recovery.
That division is deliberate. Reusing Pi keeps SEPT focused on task ownership and continuity. Each worker has one resident task slot, including while waiting. Task state, operator messages and lifecycle events are persisted, so a restart can recover the existing run and later input can wake a waiting task.
Context rollover follows the same approach. Provider-reported usage and the active model’s context window determine when a handover is needed. SEPT persists a structured handover and a human-readable record before opening a fresh conversation, carrying forward completed work, decisions, uncertainty and the next action.
The worker can be embedded through its library, controlled through an HTTP client or run from the CLI. The library and CLI share the same worker composition. Docker execution separates the agent runtime and credentials from the environment where tools act, with permissions controlled explicitly by the operator.
SEPT provides a foundation to build on, whether through direct integration into applications and agent workflows or further development of the open-source project itself. Pull requests are welcome for review and potential inclusion.
The goal was to save engineering time by addressing a bottleneck I had encountered when integrating AI programmatically into applications and agent workflows.
Install with:
npm install sept-agent
See the documentation for further guidance on integration.
SEPT is intended to run on Linux servers, where persistent workers suit the nature of long-running agent tasks. A macOS execution path is also available through Docker Desktop, with verification currently limited to specific Apple Silicon configurations.