smithersai/smithers

What it solves

Smithers provides a durable runtime for coding agents to perform complex, multi-step work on real repositories. It solves the problem of ephemeral agent sessions that are lost when a terminal crashes or a session ends, allowing long-running tasks to be persisted, inspected, and recovered.

How it works

Smithers uses a JSX-based workflow system where every completed step is immediately persisted to a SQLite database. This allows the runtime to re-render the workflow from a persisted state to determine the next task, enabling features like resuming after a crash, rewinding to an earlier frame, and forking alternate timelines. Workflows are typically authored by the coding agents themselves via plain-English prompts rather than being written by hand.

Who it’s for

Developers using coding agents (such as Claude Code, Codex, Cursor, or Pi) who need to perform substantial engineering work—like implementing features across many steps or running review loops—that requires human approval, durability, and observability.

Highlights

• Zero-config authoring: Agents build the workflows based on plain-English descriptions of the desired outcome.
• Time travel: Ability to rewind, fork, or replay runs from any specific point in the execution history.
• Durable execution: Runs survive crashes and restarts by persisting every completed step.
• ** durability across agents**: Support for multiple agents and models (Claude, Gemini, Pi, etc.) within a single workflow.
• Human-in-the-loop: Built-in gating for risky steps via human approvals.
• Cross-run memory: Durable memory that allows agents to recall information from previous runs.
• Observability: Includes a dedicated monitor for live tracking, Prometheus metrics, and OpenTelemetry traces.