I Tried Running Four Agents in Parallel. Twenty Minutes of Silence Is Where It Broke.

You spawn four agents and three push branches on schedule. The fourth has been silent for twelve minutes, and you will not notice until you check.

I learned this the hard way. I was running a Claude Code agent team, four sessions, each on its own task. Mid-sprint, one hit the Anthropic quota ceiling and stalled. The other three kept going. I did not find out for twenty minutes.

Switching the stalled agent to a fallback model got the task unstuck. It did not fix the bigger problem. I had no way to know which agents were healthy and which were drifting, without checking each one by hand.

What I actually changed

Not the agents. The inspection.

Each agent now runs in its own Herdr workspace, its own worktree, its own pane. The orchestrator session sits outside all of them. When one feels quiet, I do not attach and interrupt. I read its last buffer from outside.

herdr agent list
herdr agent read <name>
herdr agent get <name>

agent list gives the fleet state in one table. agent read dumps the last terminal buffer without switching panes. agent get returns structured status: workspace, pane, interactive state, agent kind.

When the buffer shows a loop or a stall, I stop the agent from the orchestrator with herdr agent send-keys. When it shows steady progress, I let it run.

What changed in practice

On my machine, the sustainable ceiling is three heavy worktrees in parallel. Four is possible but I lose track of one inside ten minutes. Two is comfortable but wastes the setup.

The number is observation, not contract. Yours depends on machine, model quota, and how often you can pause to inspect. The point is the ceiling is set by your observation budget, not by your spawn capacity.

The move I use most: pause the orchestrator, run herdr agent list, read each agent that has been quiet for more than a few minutes. If one has drifted, I catch it in the buffer before it produces a PR I will reject.

What worktree isolation does not fix

Herdr gives you physical isolation. Each worktree has its own files, its own branch, its own session. It does not give you logical isolation.

If agent A is changing the auth schema and agent B is building on the old one, both worktrees build clean. Both PRs pass review in isolation. The merge still breaks. Observability catches the silent agent. It does not catch the silent dependency.

That part is on you, at the planning layer.

Before you spawn four

Write the one command you will run to catch the silent agent. If that line is blank, you are not ready to spawn four. You are ready to lose an hour.