Comparison
Bystep vs AO
Agent Orchestrator (AO) by Untrivial-ai: an open-source “Agent IDE” — an Electron app over a local Go daemon — for supervising a fleet of coding agents (27 harnesses, including Claude Code, Codex and Cursor), each in its own git worktree. Failed CI, review comments and merge conflicts are routed back to the agent that owns the branch.
What AO is
Agent Orchestrator (AO) by Untrivial-ai: an open-source “Agent IDE” — an Electron app over a local Go daemon — for supervising a fleet of coding agents (27 harnesses, including Claude Code, Codex and Cursor), each in its own git worktree. Failed CI, review comments and merge conflicts are routed back to the agent that owns the branch.
- Pricing
- The app is free and open source (Apache-2.0). Paid design-partner pilots run US$500–2,000/month; AO Cloud has no public price — per useao.dev and its repository when we checked on 10 Sep 2026.
Competitor facts checked 10 Sep 2026 · Sources: useao.dev, github.com/Untrivial-ai/agent-orchestrator
Feature matrix
Where the two actually differ
Every Bystep item here is live in the product today — not a roadmap promise. Competitor claims we couldn't verify first-hand are marked “as far as we know.”
| Bystep | AO | |
|---|---|---|
| PRD generation | Yes | NoAn orchestrator agent can write a plan in chat, but there is no PRD step |
| Acceptance criteria per task | Yesids like AC-checkout-1, derived from the PRD, editable by hand | NoNo structured acceptance-criteria field per task, as far as we know |
| Verification gate | Yes`task complete` returns 409 without passing evidence or a recorded override | PartialAn optional reviewer agent and the PR/CI loop come before merge, and merging waits for you; nothing requires evidence for the task itself |
| Evidence panel | Yestest output, diff stat, files touched, optional screenshot — stored, not just claimed | PartialPR, CI checks and review verdicts on each Kanban card — built for engineers, not a proof page for a client |
| Works with any coding agent | YesClaude Code, Cursor, Codex, Copilot, OpenCode, Gemini CLI via CLI + MCP | Yes27 harnesses; reuses the agent CLIs you installed and their existing logins |
| Import / export formats | YesSpec Kit, OpenSpec, Backlog.md, Task Master, Kiro, Claude Code Tasks, GitHub Issues, Linear | NoCan turn GitHub/GitLab issues into sessions, but is not a spec/task format hub |
| Client mode / proposal | Yesclient-viewer share page + proposal for the Studio tier | No |
| Cost per feature | Yesper feature, not just per session | PartialEstimated USD per session and Kanban card from native transcripts and a pinned price catalog — per session, not per feature |
| Indonesian language & payments | Yesen/id UI, IDR pricing, QRIS/VA | NoEnglish product |
| Open-source CLI | Yes`@bystep/cli` + `@bystep/mcp`, MIT | YesThe whole app and daemon, Apache-2.0 |
Honest answer
When to choose AO instead
- Your team lives in GitHub/GitLab pull requests and wants agents to fix failing CI, answer review comments and resolve merge conflicts on their own — this is where AO goes furthest.
- You want a fully open-source desktop app for macOS, Windows or Linux, with a phone companion over your LAN or Tailscale, and your code never leaving your machine.
- You are comfortable letting the orchestrator agent decide priorities — AO has no task dependency graph or concurrency limit. Bystep instead runs a fixed plan with dependencies and requires evidence before a task counts as done.
Keep your agents, add the proof
Bystep runs the agents you already use from one plan, then requires evidence before a task counts as done. The result is readable by people who don’t read diffs.