Bary
Run AI coding agents from different vendors as one team — with enforced cross-vendor review, pass/fail gates, and a budget cap that stops runaway spend before it happens.
Running many agents isn't the same as a team that checks itself
Desktop orchestrators already let you run a dozen agents in parallel worktrees. What's missing is accountability: nothing stops an agent from approving its own patch, nothing caps spend before it's too late, and review still falls back on you.
Self-review
One agent writes the code; nothing independent checks it before merge.
Cost blowups
Fan-out with no spend ceiling has produced real four-figure bills overnight.
Review burden
"My time shifted from producing code to reviewing code" — a common complaint once agents multiply.
Plan → implement → cross-vendor review → test → merge
Describe the pipeline once
A small YAML file lists the roles: who implements, who reviews, what must pass, and what the budget is.
bary runs it in an isolated worktree
Each run gets its own git branch, so parallel runs never collide or overwrite each other.
Review is enforced, not requested
must_differ_from makes a same-vendor review a validation error — bary won't even start the run.
Gates decide if it lands
Tests pass, review quorum met, optional human approval — only then does the run pass.
Every call is on the ledger
Cost per agent, per step, per run — logged to .bary/ledger.jsonl and checked against the cap before every call.
Bary is the engine, not another desktop app
Orca, Paradev and similar tools run agents side by side well. Bary is a headless CLI and GitHub check that adds the layer none of them enforce.
| Capability | Orca / Paradev | Buzz | Bary |
|---|---|---|---|
| Run many agents in parallel | Yes | Yes | Yes |
| Cross-vendor review enforced | No | No | Yes |
| Pass/fail gates (tests, quorum, human) | No | Partial | Yes |
| Budget cap + per-agent cost ledger | No | No | Yes |
| Works inside your existing GitHub/CI | Partial (desktop app) | New workspace | Yes |
| Built on open standards (A2A/AgentBridge) | No | Uses Nostr | Yes |
Five commands
cargo install --path crates/bary-cli
cd your-repo
bary init # writes .bary/pipeline.yaml
bary validate # rejects same-vendor review
bary run --task "add a login page" --mock # try it with no agents, no cost
bary ledger # spend per agent
Read the design doc and roadmap
The full landscape research, differentiators and MVP plan live on GitHub.