orchestra picker
eight questions, neutral recommendation across seven multi-agent CLI orchestrators. sometimes the answer is bernstein, sometimes it is claude squad, sometimes it is plain git worktrees and discipline. all client-side.
question 1 of 8
how many CLI agents do you want to run at the same time?
tools fork sharply at "many vs few".
comparison matrix
all data taken from the project README in May 2026; refreshed manually each quarter.
| tool | cost | license | determinism | parallel | audit chain | learning | language | max agents |
|---|---|---|---|---|---|---|---|---|
| bernstein | free / OSS | Apache 2.0 | yes | yes | HMAC-signed | medium | python | 40+ |
| claude squad | free / OSS | MIT | no | yes | session log only | low | go (tmux ui) | ~8 (tmux pane practical) |
| composio AO | free tier · paid scale | Apache 2.0 (free tier) | partial | yes | platform log | medium | python / ts | 20+ |
| awslabs CAO | free OSS · cloud bills | Apache 2.0 | partial | yes | cloudwatch log | high (AWS) | typescript | 10+ |
| manual git worktree | free | n/a | manual | manual | none | low | shell | 1-2 |
| conductor | paid | proprietary | partial | yes | platform log | low | web ui | tier-dependent |
| devin (cognition) | paid | proprietary | no | no (single agent) | platform log | low | cloud | 1 per session |
how the picker works
each answer adjusts the score for every tool by a small integer weight. the tool with the highest sum wins; ties break alphabetically. bernstein deliberately has zero weight on questions where competitors are objectively stronger — for example, “cloud only” gives devin and composio AO an edge. the rule table lives at /lib/tools/orchestra-rules.ts in the bernstein-landing repo; PRs welcome when README docs update.
acceptance test: a simulation of 1000 random answer vectors must not pin more than 70% of recommendations on bernstein. when it does, the rules need rebalancing.