Workflows & Nuggets

Nuggetz is a shared workbench for people who use AI together. You break work into small, ownable units called Nuggets, group them into a visible workflow, and let every owner bring the agent they already use.

What a nugget is

A Nugget is one small unit of work that can be owned, handed off, run, and reviewed. Many Nuggets repeat, but repetition is not what makes them useful. Every Nugget has three things:

• An owner: the person accountable for it. Even when an agent runs the work, a human still owns the Nugget.
• A trigger: what kicks it off. A schedule ("every Monday"), an event ("a new lead lands"), or the step before it finishing.
• A definition of done (DoD): how everyone knows the Nugget is complete. Write it as a plain, checkable condition, like "draft is written and saved to the content doc."

The DoD does double duty. For a person, it is the finish line. For an agent, it is the stop condition.

A workflow is the shared board where those Nuggets add up to a real outcome, like publishing a weekly article or turning a lead into a booked meeting. Everyone can see the sequence, owners, results, and where human judgment is waiting.

When someone chooses Run with…, the Nugget stays in that shared workflow. The person remains its owner; Claude, Codex, Cursor, or another agent becomes the executor for that run. The result, any request for help, and every accepted improvement return to that Nugget's history.

The three stages

Every nugget sits at one of three stages. This is the core idea in Nuggetz: you don't flip a switch from "human" to "robot," you move one step at a time.

An Autonomous Nugget is a workflow step an agent can run end to end. It picks up the trigger, does the work, checks its result against the DoD, and reports the run. If it cannot finish, it hands the run to a person instead of guessing (see Connecting your agents).

How Nugget Chef proposes workflows

You don't have to design a workflow from scratch. Nugget Chef takes a plain-language description of what you're trying to do and proposes a set of nuggets for it.

Describe the work the way you'd explain it to a new teammate, for example:

Every week we publish one article. We pick a topic, write a draft, edit it, and publish it.

Chef turns that into proposed nuggets, each with a suggested owner, trigger, and definition of done. You then:

• Accept the ones that look right
• Edit any that need adjusting (rename, rewrite the DoD, change the owner)
• Delete the ones you don't need

Nothing is locked in. Chef gives you a starting point; you shape it.

Edit the flow after you create it

Open a workflow and use the controls beside its steps. The flow is always ready to edit, so there is no separate editing mode:

• Add step here inserts a new sequential step between two existing steps. It runs after the step above it and before the step below it. Nuggetz moves the later steps down automatically; their nuggets and settings stay intact.
• Add in parallel adds another nugget to the same step. The nuggets in that step can run at the same time, and the workflow waits for the whole step to finish before continuing.
• Add next step adds a new sequential step after the final step in the workflow.

A decision and its branches form one connected boundary. Nuggetz won't insert a sequential step between the decision and its branch group, because that would break the routing. Add the new step before the decision or after the branches instead. To add another branch, use Add branch on the branch step.

You can open any nugget at any time to change its owner, trigger, definition of done, or stage.

Every saved configuration edit creates a new Nugget version. History shows the changed fields as before/after values alongside agent reports, corrections, and proposals. Future runs always receive the current version, while the team can still see how the finish line evolved.

The start options describe what Nuggetz will actually do:

• By a person waits for a teammate to start the nugget from Nuggetz.
• After the previous step is for steps after the first and starts automatically when all work in the preceding workflow step finishes.
• On a schedule uses the UTC schedule picker described below.
• From another app waits for an external service to call the nugget's secure webhook. Save the nugget first; then a team admin can create and copy the trigger from its details.

For a scheduled trigger, choose Every day, Weekdays, Every week, or Every month, then select the UTC time. Weekly schedules let you select one or more weekdays; monthly schedules let you choose the day of the month. Nuggetz shows a plain-language preview before you save.

Automatic safety when agents stall or fail

Autonomous work never stays silently claimed by an agent that has stopped reporting. If an agent stalls, Nuggetz releases the work so another agent can pick it up.

Repeated failed or blocked runs pause the nugget automatically. New triggers stop until the owner reviews what went wrong and resumes it. Sending a run to a person for review does not count as a failure; asking for help is a safe outcome.

Promoting and demoting

Nuggets start at Manual. As you build trust that a step runs cleanly, you move it up.

• Promoting a nugget to Autonomous is always a human decision. Nuggetz never promotes on its own.
• After a nugget runs cleanly several times in a row, it becomes a "ready to promote" candidate, a suggestion, not an action. You still make the call.
• The owner can demote any nugget at any time. If an Autonomous nugget starts producing work you don't trust, drop it back to Assisted or Manual and it stops running on its own immediately.

Promotion is reversible and low-risk by design. You're never stuck with an automation that's misbehaving.

Autonomy coverage

The Workflows page tracks autonomy coverage, the share of your nuggets that run Autonomous.

It's measured across nuggets that can be automated, so review steps (below) don't count against you. As you promote more steps, coverage goes up. It's a simple way to see how much of your repeatable work now runs on its own.

Review steps that stay Manual on purpose

Not every nugget should become Autonomous. Some steps exist specifically so a person looks at the work: a final sign-off, a sensitive decision, a quality gate before something goes public.

These are review steps, and keeping them Manual is the right call, not a gap. Autonomy coverage leaves them out of the math so you're never pushed to automate a step that's meant to have a human in it.

A small example

Here's a "Weekly Content Marketing Engine" workflow partway through its journey:

Over time, "Draft the article" might move from Assisted to Autonomous once you trust the agent's drafts. "Edit for tone and length" stays Manual because it's a review step. That mix is normal and healthy.

Next steps

• Connect your agents so your team can run Assisted and Autonomous Nuggets
• Getting started for a walk through your first workflow