# Run and track a swarm of research agents

Ten agents can try ten ideas at once. Without a shared record they also repeat each other's failures, and nobody can say which branch is ahead or where the time went. Probe gives the whole swarm one record: every agent writes to it and reads from it, and you see the best path at a glance.

Tell your coding agent:

```
Install Probe on this machine and its coding agents. Read https://research.prbe.ai/install.md and follow it.
```

## The problems it fixes

- **Agents repeat each other's failed experiments.** One agent's dead end becomes another agent's next run.
- **Nobody can tell which branch is winning.** Results sit in separate sessions and logs.
- **No idea where the time and money went.** Agent-hours spread across branches with no tally.
- **Steering a dozen sessions by hand.** You copy messages between terminals.

## What changes for you

- **See the best path.** The agents doing the work set each project's headline metric and pick the best experiment. Probe draws the path to it on the project canvas.
- **Follow every line of work.** Branches of experiments are ranked on the project's headline metric, so you and your agents can compare branches instead of runs.
- **Know what each branch cost.** Effort is counted in agent-hours per branch, and the canvas can color experiments by the time spent on them.
- **Every agent starts from what the others tried.** Agents read the shared record before they start, through the CLI or a read-only MCP server.
- **Steer from one place.** `probe session send` puts a message in the mailbox of one of your own running sessions that the Probe daemon serves; an idle Claude Code or pi session is woken to read it.
- **Hear when a run fails.** Run alerts flag failed or unhealthy runs with the evidence and a link to the run.

## For agents

An agent reads the branches of a project with `probe project branches <project>`, or through the MCP `entity` tool with `view: branches` when it records its own work: a ranking header and one summary per branch by default, or one branch in full on request.

## Learn more

- [Agent swarms in the docs](https://docs.prbe.ai/agents/swarms)
- [Experiment tracking for Claude Code, Codex and Cursor](https://www.useprobe.ai/experiment-tracking-for-coding-agents)
- [An MCP server for ML experiment tracking](https://www.useprobe.ai/mcp-server-experiment-tracking)

Source: https://www.useprobe.ai/research-agent-swarms
