Ballet and the alternatives

Researched:

2026-10-08

AI coding tools are converging on one idea — agents doing whole tickets — but they differ in what they take responsibility for. This page compares Ballet with the three groups of tools teams most often consider, so you can tell when Ballet is the right choice and when it is not.

The landscape

Hosted coding agents

Devin, the GitHub Copilot coding agent, OpenAI Codex (cloud), Cursor’s background agents. You assign an issue; the vendor runs one agent in its cloud sandbox and opens a pull request for a human to review. Strong single agents, easy to start; the agent, the model, the sandbox and the pricing are the vendor’s. Copilot’s agent works in a GitHub Actions environment and pushes to a draft pull request; Devin keeps playbooks and a knowledge base of repository conventions; Codex runs tasks in parallel cloud sandboxes.

Parallel-session managers

Conductor, Claude Squad, Vibe Kanban, Nimbalyst, Emdash. Desktop, terminal or local web apps that run several CLI agents (Claude Code, Codex, Gemini CLI, opencode, …) side by side, each in its own git worktree, with a board or session list and a diff view. A developer supervises them interactively; they do not plan, review or merge on their own, and they run where the developer’s machine runs.

Open-source orchestrators

Composio’s Agent Orchestrator, Bernstein, Baton, Microsoft Conductor. Frameworks that dispatch issues to agents in worktrees and follow pull requests and CI; some add an LLM planner that produces a task graph (Bernstein) or pre-merge checks. Mostly single-host, with tracker integration through GitHub issues.

What Ballet does differently

Ballet is an orchestrator of process, not another agent. It runs existing agents — Claude Code and opencode today — through the steps a good team follows, and keeps humans in control of the plan.

Plans are approved, not improvised.

A planner agent turns a conversation into milestones, epics and tickets with acceptance criteria and dependencies, proposed as a changeset a human approves (Workflow). Execution follows the dependency graph, in parallel where it can.

No agent grades its own work.

Implement, review, verify and integrate are separate sessions in fresh workspaces that share artifacts, never transcripts (Ticket Pipeline). Hosted agents and session managers leave review to the human; Ballet makes it a pipeline stage and keeps human review as a per-ticket option.

Questions do not need a babysitter.

Agents ask only what they must; the planner answers from the knowledge base when it can; the rest waits in one inbox. An answer given within minutes goes into the running session; later, the parked session resumes with it. Only the affected ticket waits (Questions and Human Involvement).

Knowledge grows with every ticket.

Documentation, decisions and debt are written back to a per-organization knowledge base, and every human answer becomes a decision record (Knowledge). Each new session starts from that onboarding bundle instead of a blank context.

Your agents, your models, your infrastructure.

Ballet is self-hosted. Sessions run on a fleet of agents built from each project’s devcontainer — a laptop, VMs or N containers on Kubernetes — with no Docker socket or cluster API in Ballet (ADR-0025: Agent Fleet Runs Sessions as Processes). Every model call goes through Ballet’s gateway with the organization’s own key.

Safe to leave alone.

Token budgets per ticket, project and organization, enforced at the gateway; pause and kill switches; retries and stuck detection; durable state; a morning digest (Unattended Operation).

Built for many organizations.

Organizations are hard isolation boundaries for code, knowledge, keys and budgets, with OIDC sign-in and role bindings (Security) — one platform serves many teams and companies, not one developer’s afternoon.

At a glance

Typical of each group as of the date above; individual products vary.

Ballet

Hosted agents

Session managers

OSS orchestrators

Plan from a conversation, approved by a human

Yes

No

No

Some

Dependency-ordered parallel execution

Yes

No

Manual

Some

Separate review and verify sessions

Yes

No

No

Some (CI and checks)

Questions routed, answered live or resumed

Yes

Chat with the agent

Interactive only

No

Knowledge written back per ticket

Yes

Vendor knowledge features

No

No

Choice of agent runtime

Claude Code, opencode

The vendor’s

Many CLIs

Many CLIs

Runs unattended on your servers

Yes

Vendor cloud

Developer’s machine

Usually one host

Budgets and kill switch

Yes

Plan limits

No

Some

Multi-organization isolation

Yes

Teams and organizations

No

No

When to choose something else

  • You want to pair with one agent in your editor — use the agent directly (Claude Code, Cursor, Copilot). Ballet is for work you hand off.

  • You want results today without running anything — a hosted agent is quicker to start; Ballet must be installed and given agents with your toolchain.

  • You need a runtime or forge Ballet lacks — Ballet drives Claude Code and opencode and has a GitHub adapter (other git hosts work without pull-request automation); session managers support many more CLIs.

  • Ballet is young. Its first version targets an unattended night of work on a real project; the Gantt view, more runtimes and forges and notification channels come later (Vision).

Sources