AI Workflows

AI Workflows (in the sidebar) are your workspace's processes, written down so AI agents follow them instead of improvising. A workflow says, step by step, how a kind of work should be done — how a task gets planned, how it gets implemented, how a new task should be created — and every agent that works in your workspace runs that same process: your local AI client connected over MCP (e.g. Claude Code) and the built-in agents you @mention in discussions alike.

Define the process once, and it stops depending on who — or which AI — happens to be doing the work.

What a workflow is

A workflow is a named, editable document with structure:

Part What it is
Name & description The description is what an agent reads to decide when the workflow applies — write it like an instruction to a colleague ("Use whenever the user wants to plan a task by key…").
Four phases Preprocessing → Processing → Validation → Postprocessing. Gather context, do the work, check the work, record the results. Because finalization comes after Validation by structure, an agent can't "forget" to verify.
Steps Each phase holds ordered steps — a title plus Markdown instructions. Steps can be switched off without deleting them, so you can tune a process while keeping the option to bring a step back.
Arguments Inputs the workflow needs (for example a task key like BE-51), described so the agent knows what to pass.

The three built-in workflows

Every workspace starts with three, covering the core task loop:

  • task_plan — turn a tracker task into a reviewable plan: read the task, the relevant docs and learnings, fill in the Task Plan template, validate it for completeness, and attach it to the task. Planning never writes code.
  • task_implement — execute an approved plan: implement, run the tests (mandatory, in the Validation phase), then record a result, a learning, and any reusable learnings back into the knowledge base, and commit in logical chunks.
  • task_create — your task conventions, enforced: the right team, mandatory labels, an estimate, an assignee, a priority, and a structured description with acceptance criteria — then the agent re-reads what it created and fixes anything missing.

They're a starting point, not a limit: edit them, or add workflows of your own for any recurring process — releases, incident write-ups, code review checklists, onboarding.

The built-in defaults keep improving with the product. As long as you haven't edited a default, it updates automatically; the moment you customize one, your version is yours and is never overwritten.

How agents pick them up

There's nothing to install or sync:

  • Your connected AI client sees each workflow as a tool (named wf_task_plan, wf_task_create, …) and as a slash command. Because the agent reads the workflow's description, plain language just works — say "plan BE-51" and the right workflow runs.
  • The built-in agents you @mention in discussions follow the same definitions.
  • Changes reach agents instantly. Edit a workflow and every connected client updates its tool list live — no reconnect, no restart.

Plain language is all it takes — server-side agents included

You never have to name a workflow to run one. Every workflow carries a description of when it applies, and agents match your request against it — so an ordinary sentence is a complete invocation, in whatever words (and language) you'd naturally use.

This works the same in a discussion as it does in your editor. @mention a server-side agent in any task thread, channel, or brainstorm and ask for the outcome:

@agent create a task for the login-timeout bug

The agent doesn't improvise a task from scratch — it recognizes that this is what the task_create workflow is for and follows it: resolves the right team, sets the mandatory labels, estimate, assignee, and priority, writes a structured description with acceptance criteria, then re-checks the result before replying with the task key. "Plan BE-51" in a comment runs your planning process the same way; "summarize and file the follow-ups from this thread" ends in tasks that all follow your conventions.

The practical effect: the quality of agent output stops depending on how carefully each request is phrased. A one-line ask from anyone on the team — to any agent, local or server-side — produces work shaped by the process you defined once.

Bring your own process — just describe it

You don't have to build workflows by hand. A connected agent can author them for you: describe your current process in plain words, and ask the agent to turn it into a workflow. "We always write a short design note before coding, and every bugfix needs a regression test — make that our implementation workflow" is a complete specification. The agent creates or edits the workflow; you review it on the AI Workflows page.

This is the fastest way to migrate the processes your team already has — from a wiki page, a checklist, or just how things are done — into something every agent follows automatically. See Authoring with AI in this section.

In this section

  • Customizing workflows — the constructor: editing steps and phases, per-team and per-member variants.
  • Authoring with AI — creating and editing workflows by talking to an agent.
  • Artifact templates — the documents workflows produce, and how to add your own.