Workflows

A workflow runs several agents in one session, each step with its own role, model and effort.

Workflow builder

The workflow builder for Duplicate credit fix: a goal, the Orchestrated, Custom and Preset modes, the plan preview with the Orchestrator row on GPT-5.6 Sol, guidance, Starts Now, Autorun, Spend cap and Start workflowThe workflow builder for Duplicate credit fix: a goal, the Orchestrated, Custom and Preset modes, the plan preview with the Orchestrator row on GPT-5.6 Sol, guidance, Starts Now, Autorun, Spend cap and Start workflow

Shape a run before it starts: a name, a goal and a mode, with the plan previewed as a recipe, step 1 on top, in the same nodes the run will show. Starts, when to ask and Spend cap sit under the plan.

Orchestrated

Give a goal and let a model pick each next step, with its role, provider, model, effort and a reason, until it says done or blocked. The Orchestrator row holds its own model and an optional Guidance box.

Preset and Custom

The workflow builder in Custom mode: Describe the steps, Auto and Generate plan, the Draft with planner button, an Add step row and the Save as preset switch next to Start workflowThe workflow builder in Custom mode: Describe the steps, Auto and Generate plan, the Draft with planner button, an Add step row and the Save as preset switch next to Start workflow

Write the steps yourself with Add step, describe them and press Generate plan, or press Draft with planner. Save as preset keeps what you changed.

The workflow builder in Preset mode with the Pick a preset list open: Trace and fix, Harden an endpoint and Migrate a contract, each with a description and its step dotsThe workflow builder in Preset mode with the Pick a preset list open: Trace and fix, Harden an endpoint and Migrate a contract, each with a description and its step dots

In Pick a workflow mode, pick a saved sequence from the list and edit any step before starting.

Built-in workflows

Start from Refactor, Plan and ship or Fix a bug. A built-in you delete stays deleted until you press Restore built-in workflows. Restore keeps the existing workflow and step identities. If its name is held by another workflow, or one of its runs is still active, the confirmation stays open and says what must change first.

Saved steps

The Workflows studio on the Saved steps tab: the 8 built-in steps Scout, Investigate, Plan, Implement, Test, Review, Resolve comments and Update docs, and an empty This workspace list with a New step buttonThe Workflows studio on the Saved steps tab: the 8 built-in steps Scout, Investigate, Plan, Implement, Test, Review, Resolve comments and Update docs, and an empty This workspace list with a New step button

Reuse steps across workflows. The 7 built-in steps are read-only, and Save a copy makes your own.

Model per step

The model picker open on the Orchestrator row of the workflow builder, with the provider icons and the Model, Version, Variant and Effort rows set to GPT, 5.6, Sol and HighThe model picker open on the Orchestrator row of the workflow builder, with the provider icons and the Model, Version, Variant and Effort rows set to GPT, 5.6, Sol and High

Put a light model on a scout and a strong one on the planner. A step follows its role by default; Pin a model picks its own provider, then Model, Version, Variant and Effort, and Goodboy records what actually ran.

When to ask

Choose when a run asks: Ask before each step, Ask after the plan to stop once for your Approve plan and then run on its own, or Run on its own to move to the next step without you, per run or for the whole session. A guard stops an agent after 4 unattended turns in an hour, and anything you write to it resets the count.

Pause, resume and skip

Pause lets the step in flight finish and starts nothing new, even after a restart; Resume picks up where the run was. Skip passes any step that has not finished, live or stuck, after one inline confirmation, and keeps what it already wrote. When a running step says nothing for 15 minutes, the run says so and offers Ask it to continue or Skip.

Run defaults

The Run defaults tab of Workflows sets what every new run starts with: how it asks, the spend cap, and which providers get the next step. The provider policy has one home: Run defaults shows When a provider is out with its summary and a link, Open Providers & models, where Send new steps to the provider with the most room sends steps with no pinned provider to the provider with the most 5h and weekly room, keeps a provider at its limit out of new work, and tells you in one sentence where the next step goes. It starts on in workspaces created from 0.16.0 and off in older ones, and it stays off until a provider that reports limits is on. Guidance is text every run starts with: the orchestrator reads it in orchestrated runs, and in custom and preset runs it goes to the roles that write code, Implementer and Docs, unless you pick others under Edit, and Polish tidies it into one rule per line. In a custom or preset run, a Guidance tag marks each step that received it. The builder opens filled from them. When a control differs for one run, a From your rules line names what changed and the control gets a dot and a Reset; with nothing changed, the line is not there. A run keeps the rules it started with.

Spend cap

The Spend cap popover in the workflow builder, switched on at $12, with Pause workflows and Only warn me, and the Spend cap chip reading $12.00 · PauseThe Spend cap popover in the workflow builder, switched on at $12, with Pause workflows and Only warn me, and the Spend cap chip reading $12.00 · Pause

Cap what a run or a session can spend, and choose Pause workflows or Warn only. A paused run offers Raise limit.

Chained starts

The Starts menu of the workflow builder open with Now, Manually and After Harden an endpointThe Starts menu of the workflow builder open with Now, Manually and After Harden an endpoint

Line work up behind work. Starts takes Now, Manually to keep the run queued until you start it, or After followed by a workflow name, which starts once that workflow completes.

Workflow run

The Duplicate credit fix run on HBL-412: three scouts working in parallel under step 1, then a planner, two implementers and a tester, each on its own model, with Waiting on step 1 above the tree and $0.13 spent of the $12.00 spend limitThe Duplicate credit fix run on HBL-412: three scouts working in parallel under step 1, then a planner, two implementers and a tester, each on its own model, with Waiting on step 1 above the tree and $0.13 spent of the $12.00 spend limit

Follow a run from top to bottom: the header with the status and Pause and Stop run as one pair, the steps in their own scroller with a line under the header once they scroll, and the composer at the bottom. Sub-agents sit indented under their agent, and a finished set folds into one row. While the run waits for its plan, Review plan opens it in a drawer. Add steps to a live or finished run. Parallel scouts appear as branches under their step. After a restart, one Resume all above the steps starts every stopped agent in that run. Deleting a run deletes its agents and their open questions with it.

Hints to the orchestrator

The orchestrator strip on step 4 with the Tell the orchestrator something box, Queue and Read now, a queued hint reading Replay the event from the Sentry trace before the tester signs off, and Show read (2)The orchestrator strip on step 4 with the Tell the orchestrator something box, Queue and Read now, a queued hint reading Replay the event from the Sentry trace before the tester signs off, and Show read (2)

Steer a run while it goes. Write a hint on as many lines as you need, in markdown, with images the next agent gets. Queue waits for the next decision, and Read now stops the step in flight, keeps what it wrote and decides again. A queued hint shows Waits for the next decision, and read ones show the step they were read at.

Why each step, and the run recap

The Recap of the Duplicate credit fix run, its Goal, and the Why each step list with 5 decisions, from Replay one event three times back to Trace where a retried webhook postsThe Recap of the Duplicate credit fix run, its Goal, and the Why each step list with 5 decisions, from Replay one event three times back to Trace where a retried webhook posts

See why each step was picked, with a short Recap of what is done and what is left, rewritten after each decision.

Step handoff summary

Start each step from a short brief instead of the whole previous chat. If the summary fails, Goodboy keeps the start and end of the chat and marks the brief as degraded.

Open questions

The Questions lens: 3 waiting and 1 blocking in a Waiting on you list, and the question Which queue should the delivery retries run on? from the Add the stuck-delivery banner agent, with the file src/deliveries/retry.ts and the answers Dedicated retry queue (Recommended), Shared jobs queue, Retry in process and Something else, plus Let an agent decide, Skip and AnswerThe Questions lens: 3 waiting and 1 blocking in a Waiting on you list, and the question Which queue should the delivery retries run on? from the Add the stuck-delivery banner agent, with the file src/deliveries/retry.ts and the answers Dedicated retry queue (Recommended), Shared jobs queue, Retry in process and Something else, plus Let an agent decide, Skip and Answer

Get each question as one card: who asks, whether it is Blocking, the question, and its answers as numbered tiles with the recommended one tagged. Pick with a number key, answer with Enter, or write something else in line. The Questions lens lists what waits on you next to the question you are answering, and j and k move through the list. The same card sits at the end of the transcript and at the top of the agent's Brief. The answers to one agent go out together once you answered its last question, and until then each one keeps an Undo. A blocking question holds its step until you answer, or until you write what you want to say and choose Send as message: the agent gets it as a normal message and the question closes. Each question shows once in the session activity, on the row of the agent that asked, and Answer on a workflow row opens that agent right at its question.

Let an agent decide

The Brief of the agent Reuse the token footer in the brief: a header with the title, the Brief and Transcript tabs and a delete icon, one line with Implementer, Running and Opus 5, and its blocking question open on Let an agent decide: An agent decides with Sonnet 5 · Medium, a hints box, Cancel and Hand offThe Brief of the agent Reuse the token footer in the brief: a header with the title, the Brief and Transcript tabs and a delete icon, one line with Implementer, Running and Opus 5, and its blocking question open on Let an agent decide: An agent decides with Sonnet 5 · Medium, a hints box, Cancel and Hand off

Hand a question to another agent with a hint and a model from Let an agent decide, then press Hand off. Its answer counts as yours, and Answer it yourself takes it back.

Import workflows

Bring custom workflows over from your other workspaces with Import in the Workflows studio.

Reversible workflow edits use the shared app Undo stack. The Undo toast lasts about 10 seconds, and Cmd+Z outside text fields also undoes the latest operation.