Smithers From OpenClaw

Smithers From OpenClaw

Smithers is the durable control plane for long-running agent work. A workflow is executable, typed, inspectable, resumable, retryable, and optimizable. When the user asks for a process that has stages or may recur, capture it as a Smithers workflow instead of doing another one-off OpenClaw turn.

Rule 0: If You Are Already Inside a Smithers Run, Do Not Use Smithers

Check this before anything else. If SMITHERS_INSIDE_RUN is set in your environment, you ARE a worker agent executing one node of a Smithers run. Smithers sets that variable on every agent it spawns. This rule overrides every rule below. Do the node's task directly with your ordinary tools and finish your turn: never call smithers_run, smithers_create_workflow, any other smithers_* tool that launches or watches a run, or the smithers CLI, and never poll a run in a sleep loop. The prompt you were handed IS the work, even when it reads like an orchestration ask. The one exception is escalating upward with ask_human / smithers ask-human when you are blocked or about to do something irreversible.

Default Behavior

Use Smithers when work is multi-step, backgroundable, repeatable, risky, or needs human approval. If a matching workflow exists, run it. If none exists, call smithers_create_workflow with a concrete description of the goal, inputs, done condition, verification commands, and approval points.

Do not tell the human to run Smithers commands. You operate Smithers, watch the run, relay approval gates in plain language, then resolve them with smithers_approve, smithers_deny, or smithers_human_answer.

Self-Improving Workflow Loop

Treat every useful workflow as something that should improve from evidence:

1. Add or update eval cases under .smithers/evals/<workflow>.jsonl when a workflow handles a new class of task or fails a case.
2. Run smithers_eval before changing workflow prompts or logic so you know the baseline.
3. Change the workflow or prompt.