Parallel AI agents - bead

For parallel AI agents, choose the primitive that matches the state your tasks need.

You want	Use	You manage
Independent live machines	Fleet	Commands, results, and deletion for each machine
Copies of a running process and its memory	Fork	Each child and the source
Independent command shards with automatic cleanup	Swarm	Job input, terminal status, and results

Start independent machines

bead fleet --template agents --size small --count 2 --kind temporary --timeout 900

Inspect every member result before dispatching work. Upload or clone your repository separately on each machine, then send a prompt.

Branch prepared memory

bead sandbox fork ID --count 2 --timeout 300
bead sandbox fork-get JOB

Fork captures the source once. Wait for each successful child and verify memoryRestored: true. Children keep the source size. Open network connections break and tenant secrets already in memory can be copied.

Run command shards

bead swarm run 'printf "%s\n" "$BEAD_SWARM_INDEX"' --count 2 --concurrency 2
bead swarm status JOB
bead swarm results JOB --output results.jsonl

Use a template or snapshot with your code already prepared, or clone a public repository inside the command. Check each member’s exit code. The swarm controller deletes its machines; fleets and forks require your cleanup. For machine-to-machine services, read private networking. Reachability depends on placement.