The Few Concepts You Need

The minimum mental model to talk to your agent about Smithers: durability, loops, approvals, rewind, any model, isolation.

You need six ideas. Each is one sentence, one picture, and one thing you can say to your coding agent. You don’t write the workflow; you describe the work in plain English, your agent renders it, and the runtime does the rest. These six are what make “the rest” worth having.

1. Durability: finished work stays finished

Every completed step is saved. If the machine dies mid-run, Smithers resumes from the last completed step and never re-runs work that succeeded. That’s the render → execute → persist loop: run a step, persist its output as a frame, and a crash just means reload the last frame and continue. (Deep version: How It Works.)

“Kick off the implement run. If my laptop sleeps, just resume it.”

2. It loops until true, not once-and-done

Smithers iterates. It runs implement → check → review until a condition is met, instead of taking one swing and handing you the result. One-shot output is where slop comes from; a loop with a real exit condition is how you get quality.

“Keep implementing and re-running the tests until they all pass.”

3. Approvals: a paused run is just a row in a database

You can put a human in the loop. A run waiting for approval is paused as a row in SQLite, so it costs nothing while it waits: no process burning, no timer, no idle compute. Pause for a day or a week; approve, and it resumes from exactly where it stopped.

“Plan the migration, then pause for my approval before touching the database.”

4. Time travel: rewind, fork, retry

Every step is a saved frame, so you can go backwards. Rewind to any earlier frame, fork from there, and try a different approach without losing the original run. The first attempt isn’t gone; it’s a branch you can compare against. (Details: How It Works.)

“Rewind that run to before the refactor and try a different approach.”

5. Any agent, any model

Smithers is not tied to one AI. Any agent, any model, any machine. A frontier model can plan while a cheaper one fans the work out: you pick the right brain for each step.

Step	Good fit	Why
Plan / architect	A frontier model	Hard reasoning, one call, worth the cost
Fan-out implementation	Cheaper, faster models	Many parallel tasks, each small
Review / second opinion	A different model entirely	Independent eyes catch what the author missed

“Use a frontier model to plan, then fan the implementation out across cheaper models.”

6. Isolation: parallel work doesn’t collide

When Smithers runs work in parallel, each agent gets its own worktree or sandbox, so two agents editing the same repo don’t trample each other; the results merge back cleanly. That’s what makes fan-out safe: ten tickets, ten worktrees, ten agents, no shared mutable mess. The kanban workflow does exactly this.

“Work all the open tickets at once, each in its own branch.”

The one rule above all six: you drive it through your agent

You never hand-write any of this. You describe the outcome and drive it through your agent, which renders the loop, the gate, the fan-out, the isolation. These six are the vocabulary for asking for exactly the run you want.

Read next

Talk to your agent

How-to: how the conversation actually works, what to say, what comes back.

What you can do

Reference: the built-in workflows and the kinds of work you can hand off.

How It Works

Explanation (deeper): the render, execute, persist loop, frames, and resume in detail.

Why React?

Explanation (deeper): why a JSX runtime, and why time travel comes for free.