Running parallel agents is easy until you can’t see what they’re actually doing. A new open-source plugin called claude-agent-map aims to fix the visibility crisis in multi-agent workflows by providing a dedicated, read-only visualization pane within Claude Desktop. Developed by iha2, the tool transforms the opaque 'pile of IDs' generated by swarm runs into a structured, hierarchical map that displays session relationships, task status, and resource consumption in real time.

From Grep Logs to Visual Maps

The core problem with frameworks like superlite is that while they happily launch a dozen agents simultaneously, the resulting transcript becomes a waterfall of launches and checkpoints that is difficult to parse manually. The author notes that while grepping progress.md and metrics.jsonl technically works, it lacks the immediacy of watching a live run. claude-agent-map solves this by opening a pane beside the main transcript that shows the main session at the top and every subagent nested under the specific agent that spawned it.

Granular Metrics and Status Indicators

The visualization isn’t just a static tree; it provides actionable telemetry for each node. Users can see who the agent is (type and description), what it’s working on (linked to superlite task IDs), and how it’s performing via color-coded status indicators: green for running, red for aborted or blocked, and gray for completed tasks. The pane also tracks cost metrics, including tier, elapsed time, and token usage, with a detail view available for model specs, parent relationships, and the last message sent.

Integration and Workflow

Installation is streamlined for superlite users, as the plugin is pulled in as a dependency from the marketplace. It can be invoked directly via the /claude-agent-map command, pointed at a specific project folder, or triggered by replying 'approve map' during plan approval. Crucially, the pane is strictly read-only; it never steers the run, ensuring that observation doesn’t introduce side effects into the agent’s execution logic. The code is available on GitHub for those looking to inspect the implementation details.

Key Takeaways

  • claude-agent-map provides a hierarchical, read-only view of parallel agent runs in Claude Desktop.
  • The tool integrates with superlite workflows, displaying task IDs, token costs, and real-time status colors.
  • It addresses the 'visibility crisis' where multi-agent transcripts become unreadable waterfalls of logs.
  • The plugin is open-source and available on GitHub, with marketplace support for easy installation.

The Bottom Line

Visibility is the missing link in agentic coding; if you can't see the swarm, you can't trust the build.