Skip to content

Workflows: processes built from blocks, with a person in the decisions

AI ENGINEERING · WORKFLOWS

Inside AI Engineering, a flow chains agents, knowledge base searches, n8n automations and human approvals in a design the team builds on screen or starts from a ready-made flow. Only the published version runs, and every run is recorded step by step.

updated

01 · Lifecycle

From design to run history

An editable draft, a published version that never changes, step-by-step runs and a person wherever there is a decision.

  1. DESIGN

    The team builds the flow on screen by connecting blocks; the draft can change freely.

  2. PUBLISHING

    Publishing checks the design and saves a version that never changes again.

    immutable version
  3. START

    The flow starts from the screen, on a schedule or through the API.

  4. STEPS

    Each block runs in design order and passes its output to the next one.

  5. APPROVAL

    At the approval block, the run stops and waits for a person to decide.

    awaiting approval
  6. HISTORY

    Input, output and duration of every step stay in the run history.

    auditable

02 · Blocks

The step types

The 10 blocks in the editor. In any text field, {{entrada}} becomes the output of the previous step.

  1. Start

    Sets how the flow can start: from the screen, on a schedule or through the API. It can carry a default input text.

  2. Agent

    Runs the published version of a workspace agent along the same path as regular runs: Gateway, budget, limits, sensitive data policy, tool approval and the model rule of whoever started it. The step waits for the agent to finish.

  3. Knowledge base

    Searches the knowledge base with the permissions of whoever started the flow and passes the passages along. Requires a Knowledge subscription.

  4. Connector (automation)

    Triggers an n8n automation registered in the workspace, with parameters in its schema, once per step. Fire and forget: it does not wait for the n8n response.

  5. Human approval

    Stops the run and shows a person with approval permission the instructions and everything that arrived so far. Approve continues; reject ends; with no decision in 24 hours, the request expires.

  6. Condition

    Picks the "yes" or "no" branch by comparing the previous output (contains, equals, greater than, less than), on the full text or on a JSON field. No free-form code.

  7. Join

    Waits for the branches that reach it before continuing and passes the combined outputs along.

  8. Response (end with template)

    Ends the path with text written in the design, combining the previous output with outputs of earlier steps referenced by their identifier. That text becomes the run result.

  9. End

    Ends the path. A flow can have more than one end.

  10. Run another flow

    Runs the published version of another workspace flow, with the permissions of whoever started it, and waits for it to finish; its output moves on. The chain goes up to three levels deep.

03 · To get started

Ready-made flows

Starting designs, each with the agent of a solution. They always come in as drafts: the team reviews, adjusts and publishes; nothing runs before that.

  1. approval + optional connector

    Ticket triage

    The support agent classifies the ticket and drafts a reply. High priority asks a person to confirm before routing.

    Uses the AI for Customer Support agent
  2. approval + optional connector

    Reconciliation difference

    The reconciliation agent compares the statement and ledger entries and lists the differences. When there is a difference, the finance team approves before anything is routed.

    Uses the AI for Financial Reconciliation agent
  3. approval, no connector

    Sensitive HR question

    The people agent answers questions from the policy. A sensitive topic gets no answer on the merits: it becomes a routing request that someone in HR approves.

    Uses the AI for People (HR) agent
  4. approval, no connector

    Contract review

    The legal agent compares the draft with the authorized standards in the base. Legal reviews the differences and approves before the draft moves on.

    Uses the AI for Legal agent
  5. approval, no connector

    Campaign draft

    The marketing agent writes the draft following the brand guide. Someone on the team approves before the text leaves the flow; nothing is published.

    Uses the AI for Marketing agent
  • If the solution is not in the workspace yet, the installer explains what is missing; someone who also manages the base can install the solution at the same time.
  • An n8n connector goes in only when the person picks a registered automation. Without one, the approval is the routing point.

04 · In the editor

Build, compare and save

The editor helps you build what repeats and track what changed between versions.

Orchestrator

Agents in parallel

An editor shortcut sets up, from one block, 2 to 5 agents on the same input plus a join that collects the answers. It is not a new block: these are regular branches and a join, and each agent runs with its own budget, approval and trace. The result stays editable like the rest of the design.

  • Compare versions

    Between two published versions, the screen shows the blocks and connections that were added, removed or changed. Moving a block on the canvas does not count.

  • Duplicate

    Creates a new draft from the current design, without carrying over versions, runs or schedule.

  • Archive

    Removes the flow from the list and refuses new runs and edits; versions and runs are kept, and the flow can return to the list.

  • Readable history

    Every run shows block names as they were in the version that ran, even after the design changes.

05 · Triggers

How a workflow starts

The start block says which triggers apply. A disabled trigger is refused, even with permission.

  • From the screen

    Anyone with permission to run agents starts the flow on the platform, with an input text.

  • Scheduled

    At an interval (from 15 minutes to 7 days), every day or every week at a set time, always in UTC. It runs on behalf of whoever scheduled it; missed times do not pile up.

  • Through the API

    A company system starts the flow with a workspace API credential that has the flows:run scope, and later fetches the runs it started. Repeating the same key returns the same run.

  • From another flow

    The "Run another flow" block starts the published version of this flow on behalf of whoever started the calling flow, and waits for it to finish. The chain goes up to three levels deep.

  • On an event

    A workspace event (pull request opened, security finding, alert, incident, policy decision) starts the flow with the event as JSON input, one run per event, up to 30 per hour. It runs on behalf of whoever created the trigger; a flow event never triggers a flow. Enabled per organization.

Run through the API
POST /api/platform/flows/{flow-id}/runs
Authorization: Bearer <credential with flows:run>
Content-Type: application/json

{ "input": "Order 4815: travel reimbursement", "idempotencyKey": "order-4815" }

→ 202 { "id": "<run id>" }
GET /api/platform/flows/runs/{run-id}   → step by step

06 ·Capabilities

What comes with it

  • Immutable versions: only the published version runs, and the draft does not affect what is in use
  • Publishing is refused with a list of what is missing: no start, a cycle, a loose block, an agent with no published version or an automation outside the registry
  • Same permissions as Agents: whoever configures agents designs and schedules, whoever runs agents runs, whoever approves actions approves
  • Approval queue with the instructions and a summary of what reached the step
  • History of every run, step by step, with input, output and duration
  • Audit of creation, publishing, runs, scheduling, decisions, ready-made flow installation, duplication and archiving
  • Comparison between published versions, showing what was added, removed or changed
  • Up to 40 blocks and 80 connections per flow

07 ·Limits

What it does not do

  • The connector triggers the n8n automation and moves on: it does not wait for the response. The result arrives later, outside the flow.
  • Scheduling is always in UTC, with a minimum interval of 15 minutes.
  • Running through the API requires a credential with the flows:run scope and the API trigger enabled at the flow start.
  • An agent tool that asks for approval shows up in the Agents approval queue, not in the flows queue.
  • Each run has a cap on steps and duration, and there is a limit on runs per minute per person and per credential.
  • Archiving does not delete the schedule: it just stops firing. When the flow returns to the list, an overdue time runs once.
  • In ready-made flows, the condition depends on the agent writing the requested line (for example, PRIORIDADE: ALTA); if it does not, the flow follows the "no" branch.

In the product

Where workflows fit

Workflows are part of AI Engineering: they use the same agents, the same approvals and the same contract. Search uses Knowledge, and the sensitive data policy and model rules come from Guard.

Assisted rollout

How to buy

  1. Create your account

    Sign-up creates your company's organization and first workspace on the platform, hosted by AIUBY, with isolation per organization.

  2. Add products

    Choose the products and the tier under Products. Adding is free: the monthly fee starts the first time you use each product.

  3. Add credit and go

    Everything is prepaid: top up your balance and use it. Documents, agents, solutions, policies and flows stay with your team, with cost and audit in plain view.

  4. Count on the team

    If you prefer, AIUBY builds the solution for your area and follows up on adjustments. Engineering services are always delivered with the team.