Agent Manager

Agent Manager (Cmd+Alt+A) is where you build, configure, and run multi-agent workflows. It includes a visual Workflow Composer for drawing pipelines as node graphs, plus a dashboard for monitoring active and past runs.

Available in the open source edition. No Enterprise license required.

Workflow Composer

The Workflow Composer is a visual graph editor. Drag nodes from the palette, wire them together, configure each step, and run — the result is saved as .inverse/workflows/<id>.json.

Opening the Composer

Open Agent Manager (Cmd+Alt+A) and click New in the toolbar, or open an existing workflow from the Open dropdown.

Node Types

Building a Pipeline

1. Drag a Trigger node from the left palette onto the canvas
2. Drag Agent nodes for each step you want
3. Connect them: click and drag from an output port (right side of a node) to an input port (left side of the next node)
4. Configure each node: click a node to open the property panel on the right
5. Save with the toolbar Save button — writes to .inverse/workflows/
6. Run with the toolbar Run button

Canvas Controls

Nodes snap to a 20px grid by default.

Connecting Nodes

Drag from an output port (circle on the right edge of a node) to an input port (circle on the left edge). The connector turns green when the connection is valid, red when it isn't.

Connections are rejected if:

• They would create a cycle
• Port types are incompatible (e.g. a JSON output into a flow-only input)
• The target port already has its maximum number of incoming connections

Configuring a Node

Click any node to open its configuration in the right panel.

Agent node settings:

Trigger node settings:

Validation

Click Validate in the toolbar to check the graph before running:

• Error: agent node with no agent assigned
• Error: cycle detected
• Warning: disconnected (orphan) node
• Warning: trigger not connected to any step

Errors block Run. Warnings don't.

Triggers

Workflows can start automatically without pressing Run.

File Save

Fires when any file matching the configured glob is saved. A debounce delay (default 300ms) prevents repeated fires during rapid saves.

On Commit

Fires after a git commit. Optionally filtered by branch name regex or committed file paths.

Schedule

Fires on a repeating interval (configured in minutes). Starts counting from when the workflow is enabled.

Terminal Command

Polls a shell command on a configurable interval. Fires when the exit code matches your condition (failure by default — useful for "run the linter and fix errors automatically").

Manual

Only runs when you click Run in the Composer, call the API, or trigger it from another workflow.

Running a Workflow

Click Run in the Composer toolbar. The run panel slides up from the bottom of the canvas showing:

• Each step as a row with a status indicator
• Live output log as each agent streams its response
• Tool calls made per step (last 5 shown, expandable)
• Elapsed time per step and total
• A Cancel button to stop the run at any point

Step status indicators:

If a step fails, the error message is shown inline. The workflow stops at the failed step by default.

Agents

Agents are the building blocks of workflows. Each agent is defined in .inverse/agents/<id>.json.

Creating an Agent

1. Open Agent Manager
2. Go to the Agents tab
3. Click New Agent
4. Set: name, model (provider + model name), system instructions, allowed tools, max iterations

Built-in agents are provisioned automatically the first time you open an empty .inverse/agents/ directory. You can edit them in place.

Agent Fields

Workflow JSON Format

Workflows are plain JSON files in .inverse/workflows/. You can edit them directly.

{
  "id": "scaffold-component",
  "name": "Scaffold Component",
  "description": "Creates a new React component from a template",
  "trigger": "manual",
  "enabled": true,
  "steps": [
    {
      "id": "plan",
      "agentId": "planner",
      "role": "planner",
      "allowedTools": ["readFile", "listDirectory"],
      "maxIterations": 10
    },
    {
      "id": "build",
      "agentId": "code-writer",
      "role": "executor",
      "dependsOn": ["plan"],
      "allowedTools": ["readFile", "writeFile", "editFile"],
      "maxIterations": 20
    }
  ]
}

Steps with no dependsOn run as soon as their trigger fires. Steps with dependsOn wait until all listed steps complete, then receive those steps' outputs as context.

Trigger-specific fields

Built-in Workflow Templates

Agent Manager ships with pre-built workflow templates accessible from the Composer's New dropdown. Templates cover common patterns: code review, test generation, documentation sync, dependency updates, and more.

Templates are provisioned from builtinLibrary.ts and can be modified after loading.

Run History

Completed runs are stored in memory for the current session (up to 50 entries). The Agent Manager dashboard shows:

• Workflow name and trigger type
• Start time and duration
• Final status (done / failed / cancelled)
• Aggregated final output from the last completed step

Click any run to expand its step-by-step breakdown.