Workflows

A Workflow is a sequence of Skills with optional human-in-the-loop approve gates. It's how you compose single-step capabilities into an end-to-end process.

What it does

Every workflow is a small DAG of steps — and workflows are deterministic: the same structure every run. Open-ended agent work belongs in Missions, never in a workflow step. Four step types:

skill — invoke a Skill (typed LLM call) with interpolated input, capture output
sync — incrementally refresh named sources so downstream steps read live data
approve — pause, route to the Review Queue, resume on decision
action — call a side-effecting adapter (gmail.send_email, hubspot.update_deal, …)

State persists in workflow_run rows; each step's result lands in a JSONB column and is addressable as {{steps.<name>.output.<field>}} by downstream steps.

If the process dies mid-run, the run sits paused at its last persisted step. Resume from MCP, the dashboard, or a direct oRPC call.

Triggers

A workflow declares only trigger: {type: manual} (with an optional inputSchema for the Run form). Everything time- or event-driven lives in an Automation:

# workspace/<org>/automations/reply-followup.yaml
when: { event: prospect.reply }        # or { schedule: "0 12 * * 1-5" }
do: { workflow: discovery_followup }

Workflow-embedded schedule/event triggers from earlier releases still work but warn as deprecated on workspace:apply.

Folder shape

workspace/<org>/workflows/<slug>/
├── workflow.yaml      # steps + trigger + input schema (required)
├── evals.yaml         # end-to-end fixtures (recommended)
└── README.md          # optional — rationale, edge cases

`workflow.yaml`

slug: discovery_followup
name: Discovery Follow-up
description: Post-discovery-call automation. Summarize, draft, gate, send.
trigger: {type: manual}
inputSchema:
  type: object
  required: [transcript, prospect_name]
  properties:
    transcript: {type: string}
    prospect_name: {type: string}
steps:
  - name: summary
    type: skill
    skill: discovery_summary
    input:
      transcript: '{{input.transcript}}'
  - name: email
    type: skill
    skill: draft_followup_email
    input:
      discovery_summary: '{{steps.summary.output}}'
      prospect_name: '{{input.prospect_name}}'
  - name: review
    type: approve
    prompt: Review the drafted follow-up email before sending.
    reviews: email
  - name: send
    type: action
    action: gmail.send_email

`evals.yaml`

End-to-end fixtures — feed an input, assert on final output + intermediate step outputs. Catches regressions where one step's output shape shifts and breaks a downstream step. See Evals.

Runtime

Start a workflow → steps execute sequentially → on approve step, run pauses. Approver reviews at /dashboard/review, clicks approve/reject, optionally rates 👍/👎 with a note. Approve resumes the run from the step after; reject marks it failed. Durable by design — in-flight state lives in Postgres, not memory.

Every run is query-able by status, rating, context SHA — see the logs view.

Connection to other resources

Skills — every skill step invokes one
Agents — an Agent often kicks off a Workflow rather than running a one-shot skill
Review Queue — the UI surface for every approve step across every running workflow

Quickstart — run your first workflow end-to-end
Authoring context — the edit + apply cycle
Evals — fixtures catch breakage before prod
Feedback + logs — rating runs, measuring outcomes