Skip to main content
Smithers lets you hand your coding agent real, multi-step work and trust it to finish, surviving crashes, retries, and review loops. You already talk to a coding agent: ask for a change, it works for a few minutes, you read the result. That loop is great for small, synchronous edits. It falls apart the moment the work gets long.

​
The treadmill

A single agent in a chat window is a black box on a treadmill. It forgets what it decided ten steps ago, stalls halfway through a long job waiting for a nudge, and loses everything if the process crashes. You start over. The model isn’t the problem; the model is good. The problem is that nothing underneath it is durable. There’s no record of what finished, so there’s nothing to resume from.

​
From chat to durable background work

Smithers moves the work off the treadmill and into a durable runtime. You describe a real job and it runs in the background, for minutes or for days. Every step is persisted the moment it finishes, so the runtime always knows what’s done and what to run next. A crash becomes a resume point. A long job becomes a run you can walk away from. And it’s observable throughout, so you’re never staring at a black box.
Three-layer diagram: your prompt on top, Smithers as the durable orchestration layer in the middle, and interchangeable agents and models at the bottom

Smithers is the durable layer between you and the models, the part that doesn't change when the agent topology does.

​
You don’t write workflows. You talk to your agent.

You don’t click a GUI and you don’t write code. Install the smithers skill into your coding agent, and from then on you drive it through your agent. Say what you want in plain language; your agent picks the right workflow, starts the run, watches it, and brings back the result, or pauses to ask when it needs a decision. Smithers is the engine, your agent is the driver, you stay in the chat you already use. The TSX workflows and the CLI exist, and you can read them. But the primary path is a sentence to your agent.

​
What you get

Four guarantees a chat-window agent can’t give you, on any agent, any model, any machine.
One Smithers workflow running tasks across multiple different coding agents and models side by side

The same job runs across different agents and models without rewriting the work, proof that Smithers isn't tied to one lab or one harness.

​
The magic moment

Open your coding agent and say:
“Add rate limiting to the API and don’t stop until it builds, validates, and the review signs off.”
A durable run writes the code, runs the build and tests, loops back to fix what fails, and finishes only once the review approves, pausing for your sign-off at the gate. A few more things you can just say:
“Keep working on reducing our flaky tests and don’t stop until I tell you to.”
A long-horizon run that grinds on test flakiness in the background and keeps going across crashes until you call it.
“Migrate the dashboard to the new gateway APIs. Show me the milestone plan before you start, then work through it.”
Plans the migration, pauses for your approval, then executes milestone by milestone, resuming exactly where it left off if anything interrupts it.
A Smithers run is killed partway through, then resumes from the last persisted step instead of starting over

A run survives a hard crash mid-task and picks up exactly where it left off, with no lost work and no restart from scratch.

None of these need a workflow file or a terminal. You ask, Smithers runs it durably, and you get higher-quality output than a one-shot prompt, with the structure real work demands.

Get started

Tutorial: install the skill and run your first durable job through your agent.

What you can do

Reference: the outcome catalog of real jobs you can hand off, and what to say to get them.

The few concepts you need

Explanation: the handful of ideas (runs, frames, approvals) behind everything above.