Long AI workflows fail in boring ways. A step reads a file that moved, writes its result somewhere the next step doesn't look, or names a tool that isn't installed on this Mac. Until now, a failure like that meant stopping, editing the routine and starting the run again from the top. Toone 1.0.81 adds auto-fix for AI workflows and routines: when a step fails, Toone can work out what went wrong, fix the routine, and continue the same run from the step that failed.

How much Toone does on its own is your choice, per routine.

Four modes, from hands-off to hands-on

Open a routine and find Auto-fix in its inspector. There are four settings:

Off. Toone never proposes a fix by itself. When a step fails, its agent explains why, and fixes happen when you ask.
Suggest (the default). When a step fails, Toone and the step's agent propose a fix. Nothing changes until you confirm it.
Auto safe fix. Toone fixes the failed step's own text and continues without asking. The issue card shows exactly what changed, with a Revert button.
Auto-fix. Toone works out the fix, including changes to the routine's structure, and continues. It still pauses for anything only you can unblock.

Every routine starts on Suggest, so nothing changes for existing routines until you choose a different mode.

Fix and continue, in the same run

Whichever mode you pick, a fix no longer means starting over. The run keeps the work its earlier steps already finished, applies the fix, and retries the step that failed. The run's thread records what happened in plain language: what failed, what was changed, and whether the step passed afterwards.

If you would rather fix it yourself, you can tell the step's agent in the thread what to change, retry the step as it is, or cancel the run.

What auto-fix will not do on its own

An agent that rewrites your routine needs clear limits. Even on Auto-fix, Toone stops and hands the decision back to you when:

You need to act. A sign-in, an approval, a missing input or a billing problem can't be fixed by editing text, so the run waits for you.
Something may already have left Toone. If the failed step may have sent a message or changed something in another service, Toone asks you to check what was sent before anything resumes.
You may be editing. If a draft of the routine is open, or another run of it is active, Toone leaves the routine alone.
The run is on an older version. If you published a newer version since the run started, Toone doesn't patch the old one behind your back.
The routine just changed. If the routine was published in the last 30 minutes, by you or by an earlier auto-fix, Toone suggests the fix instead of applying it.

Some things an automatic fix never touches, in any mode: who a step sends to, permissions and approval steps, limits and quotas, credentials, and which provider or model a step uses. If the fix needs one of those, Toone proposes it and waits for you.

Temporary problems, such as a provider timing out, aren't treated as routine bugs. Toone retries the unchanged step twice, with a short wait between attempts, and then asks you. If a fix doesn't hold, the card says so and offers to revert it.

When Toone declines to fix something automatically, it tells you why in one sentence, for example: "the step may already have sent something outside Toone, so it is left for you."

Also in this release

One way to steer a run. Pause, continue, stop and talking to an agent now work the same way in every routine thread.
Routine bundles on your Mac. Group related routines in the new Bundles tab, then rename, reorder or remove them.
Reasoning effort for Claude. Pick how hard Claude models think, and choose Claude Opus 5.5 and Sonnet 5.5.
Updates install themselves. While you're signed in, Toone downloads new versions in the background and installs them when no routine is running. You can turn this off under Updates in Settings > General.
Safer browser sign-ins. Toone asks before importing cookies from Safari or Chrome.

Toone joins Claude Startups

We've been accepted into Claude Startups, Anthropic's program for early-stage companies building with Claude. It gives us API credits and support from Anthropic while we build, and many Toone routines already run on Claude.

This is the first of a few announcements we have lined up as we prepare for our App Store launch 👀 More to come soon.

Try it

Update Toone, open a routine you run often and set Auto-fix to the mode you're comfortable with. Suggest is a good place to start: you'll see the fixes Toone would make before you let it make them. New to routines? Start with how routines work, or pick one from Explore.