Claude: Visualize Agents Swarms news update

Superlite will happily put a dozen agents to work at once. That is the point, and it is also the problem. The author of contextfirst.blog's report split out claude-agent-map, a small Claude Desktop plugin that opens a read-only Agent map pane beside the transcript. The main session sits at the top, and every subagent sits under the agent that spawned it. The repository is iha2/claude-agent-map. The pane only reads. It never steers a run.

This piece is for people running a Superlite swarm in Claude Desktop who need the stalled agent, the task it owned, and whether the coordinator is still thinking or just waiting. It covers each node's type, Superlite task ids, green, red, and grey status, and cost in tier, elapsed time, and tokens, plus the detail view, /claude-agent-map, the reply approve map at plan approval, and the marketplace install that pulls the plugin in as a dependency.

Visualize Agents Swarms news update: what actually changed

When the build is going well, the transcript is a waterfall of launches, reviews, and checkpoints. When it is not, the watcher needs which agent stalled, which task it owned, and whether the coordinator is still thinking or just waiting. Grepping progress.md and metrics.jsonl works. The post says that search does not feel like watching the run. Superlite will put a dozen agents to work at once. claude-agent-map is what the author split out so the same run can be watched from a pane beside the transcript.

Visualize Agents Swarms news update: how it works

Each node carries four facts. Who it is appears as a type and a short description. What it is working on appears as Superlite task ids when those records are present, with a check mark once the task is checkpointed. How it is doing is a color: green while running, red if aborted or blocked, and grey when done. What it cost so far is the tier, the elapsed time, and the tokens. The main session and each subagent under the agent that spawned it use the same four facts.

The detail view adds the model, the parent, the ledger lines, and the last message. Open the read-only Agent map pane with /claude-agent-map, optionally pointing at a project or at that project's .superlite folder. Asking Claude to open it is the second path. When Superlite offers the pane at plan approval, the accepting reply is approve map. On every path the pane only reads, and it never steers a run. Tier, elapsed time, and tokens stay on the cost line.

Visualize Agents Swarms news update: why it matters now

Parallel agent workflows fail quietly in the UI long before they fail in the tests. Launches, reviews, and checkpoints can keep filling the transcript while an agent has stalled, while it still owns a task, or while the coordinator is waiting instead of thinking. Grepping progress.md and metrics.jsonl works. It does not feel like watching the run. "Eighteen agents is not visibility," the post says. "It is a pile of IDs." A dozen agents is the load Superlite will put to work at once.

The map turns that pile of IDs into a picture: who spawned whom, which tasks are still open, and which nodes went red. Red means aborted or blocked, grey means done, and green means still running. A check mark appears once a Superlite task is checkpointed. The post calls the layout just enough structure to leave a swarm running and still answer "what is happening right now?" without opening six files. That picture sits in the Agent map pane beside the transcript, and the pane only reads.

Visualize Agents Swarms news update: who is affected

People running Superlite at swarm scale are the ones in front of this pane. The post says Superlite will happily put a dozen agents to work at once, and it treats eighteen agents as a pile of IDs rather than visibility. Those runs are where someone has to name the stalled agent, the task it owned, and whether the coordinator is still thinking or just waiting. They are the runs where grepping progress.md and metrics.jsonl works without feeling like watching the run.

Installing Superlite from the marketplace pulls in the Claude Desktop plugin claude-agent-map as a dependency. The pane opens with /claude-agent-map, and the command can point at a project or at its .superlite folder. Other people can ask Claude to open the pane. At plan approval, Superlite can offer the pane, and the reply is approve map. The detail view is for the model, the parent, the ledger lines, and the last message. The node itself carries type, short description, task id, color, tier, elapsed time, and tokens.

Visualize Agents Swarms news update: what to watch

The signals on each node are the color and the check mark. Green means the agent is running, red means aborted or blocked, and grey means done. Superlite task ids show when those records are present, and the check mark arrives once the task is checkpointed. Cost on the node is tier, elapsed time, and tokens. The detail view keeps the model, the parent, the ledger lines, and the last message. Red nodes and tasks that are still open are the picture the map is there to show.

The pane only reads. It never steers a run, whether someone opens it with /claude-agent-map, asks Claude to open it, or replies approve map when Superlite offers the pane at plan approval. The command can point at a project or its .superlite folder. Installing Superlite from the marketplace pulls the plugin in as a dependency. Leave the swarm running and answer "what is happening right now?" without opening six files. These workflows fail quietly in the UI long before they fail in the tests.