Skip to main content

Workflows

In AI Company, a workflow is the map of how your agents depend on each other — who reports to whom, who calls whom, and which agents form a pipeline. You see and edit it on the Connections canvas. Clevername does not run a step-by-step workflow builder; the one executing chain is Agent Pipelines, driven through the API.

Important
This article previously described a Workflow Builder with New Workflow, Add Step, Pause/Retry/Skip failure handling and cron schedules. That UI is not part of the current product. Workflows in Clevername mean dependency mapping, described below. If you need a chain that actually executes, see Agent Pipelines.
Key Concepts

What a workflow is here

A workflow is the set of relationships between your agents: which agent triggers or feeds another, which agents an agent is allowed to call at runtime, and the reporting hierarchy. Clevername records these edges so it can govern them — drift detection compares declared peers against observed agent-to-agent traffic, and the Connections canvas shows the whole picture in one place.

The Connections canvas

AI Company → Connections (/dashboard/ai-company/connections) draws your agents, MCP servers and skills as nodes, with filter chips for Agents, MCP, Skills, Calls, Environment, Pipelines and Reporting. Edges come from three sources:

  • Reporting — the organizational hierarchy between agents.
  • Calls — agent-to-agent call edges, either declared (the upstream and downstream agents you tick when registering an agent) or observed from runtime traffic through the gateway; an edge that is both declared and observed is marked as such.
  • Pipelines — dashed green arrows from step N to step N+1 of each Agent Pipeline.

Dependency map vs. an executing chain

The map is descriptive: it tells Clevername what should talk to what, so runtime calls outside the declared edges surface as drift. It does not schedule or run anything. When you want Clevername to execute agents in order and thread output into the next step, create an Agent Pipeline through the API.

Step-by-Step Guide
1

Declare upstream and downstream agents when you register an agent

On the Register Agent page, the Agents that trigger or feed into this one (upstream) and Agents this one triggers or hands off to (downstream) lists let you declare call edges as you create the agent. You can filter the lists by team.

2

Open the Connections canvas

Go to AI Company → Connections. Use the filter chips to show or hide agents, MCP servers, skills, call edges, environment colouring, pipeline edges and the reporting hierarchy.

Connections canvas showing agent nodes with reporting, call and pipeline edges and the filter chips along the toolbar
Each chip toggles one layer of the map. Pipelines are drawn as dashed green arrows.
3

Inspect an agent

Click any agent node to open its detail panel. It lists the agent's connections and, under Pipelines, every pipeline the agent appears in.

4

Turn a chain into a pipeline

When a sequence of agents should actually run end-to-end, create an Agent Pipeline: POST /hub/agent-pipelines, add steps with /steps, then /run. Each step's on_error is halt or skip (fanout is reserved and currently behaves like halt). The full walkthrough is on Agent Pipelines.

Tip
Keep each agent in a chain focused on a single responsibility. Every hop runs through CleverGuard scanning and the Agent Review gate for that agent.