Skip to content
/
Phase B Batch 8

Batch 08 — Multi-agent Orchestrators (Parallel Workers / Swarm-style)

Batch 08 — Multi-agent Orchestrators (Parallel Workers / Swarm-style)

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
crystal 3068 desktop-app (Electron) no Electron (React) parallel-fan-out no A
maestro-runmaestro 2954 desktop-app (Electron) yes (maestro-cli) Electron (React) hierarchical yes A
maestro-its-maestro-baby 1152 desktop-app (Tauri) shell only Tauri (React) parallel-fan-out yes A
parallel-code 650 desktop-app (Electron) no Electron (SolidJS) parallel-fan-out yes A
claude-swarm 172 cli-tool (Python) yes (claude-swarm) terminal-tui (Rich) hierarchical yes (Opus→Haiku→Opus) A
swarmclaw 523 desktop-app + npm yes (swarmclaw) web-dashboard (Next.js, :3456) hierarchical yes (23+ providers) A
1code 5542 desktop-app (Electron) no Electron (React) parallel-fan-out yes A
agor 1221 npm-package yes (agor oclif) web-dashboard (React/FeathersJS) hierarchical yes A
sword-swarm 24 standalone-repo (Python) no none hierarchical unknown A
dorothy 291 desktop-app (Electron) no Electron (Next.js) hierarchical yes A

Intra-batch patterns

All 10 frameworks share the "parallel session manager" core — the ability to run multiple AI coding agents simultaneously, each on separate work. Eight of ten ship Electron, Tauri, or web desktop UIs rather than CLI tools, reflecting that multi-session management is inherently visual. Six of ten use git worktree isolation as the primary safety mechanism, though Parallel Code uniquely adds Docker container isolation and claude-swarm uses no isolation at all (relying on file locking instead). The "pass-through" philosophy dominates the desktop apps (Crystal, RunMaestro, its-maestro-baby, Parallel Code, 1Code) — they add no behavioral methodology to agents. The more opinionated systems (claude-swarm, SwarmClaw, SWORDSwarm, Dorothy, agor) inject their own orchestration protocols, memory systems, or agent hierarchies.

A clear split emerges: session managers (Crystal, its-maestro-baby, Parallel Code — worktree isolation, no behavioral modification, human picks winner) vs autonomous orchestrators (claude-swarm, SwarmClaw, Dorothy, SWORDSwarm — agents are given orchestration protocols, memory APIs, and quality gates with less human oversight). RunMaestro, 1Code, and agor sit in between, offering both interactive sessions and automation features.

Most interesting finds

  1. claude-swarm (affaan-m) — The only pure CLI tool in this batch, and the only one with JSONL session recording + CLI replay (claude-swarm replay <id>). The explicit multi-model quality gate (Opus designs and reviews, Haiku executes) is the most principled cost-optimization architecture in the corpus, backed by a clear "senior architect / junior engineer" metaphor. The --demo flag (animated TUI without API key) is an unusually clever onboarding mechanism.

  2. agor (preset-io) — The Figma-style 2D spatial canvas where branches are draggable cards that auto-trigger agent prompts when dropped into zones is completely unlike any other framework in this batch or the 11 seeds. Real-time multiplayer (live cursors, facepile, shared sessions) makes it the only team-first framework studied. Session genealogy (forked/spawned ancestry tracking) enables a novel approach to exploring alternative agent approaches.

Items written as Tier C stubs

None. All 10 frameworks had sufficient public material for full 11-file reports.

Cross-references discovered

  1. crystal → nimbalyst: Crystal is deprecated (Feb 2026) and explicitly replaced by Nimbalyst (https://github.com/Nimbalyst/nimbalyst). The two are the same product family but different repos.

  2. maestro-runmaestro vs maestro-its-maestro-baby: Completely independent implementations of the same conceptual space ("Maestro" parallel worktree manager). Not forks of each other. RunMaestro (Electron/Node, AGPL-3.0, 2954 stars) vs its-maestro-baby (Tauri/Rust, MIT, 1152 stars). The shared name appears coincidental.

  3. 1code → openspec: 1Code has an openspec/ directory, indicating use of the OpenSpec seed framework for spec management. The only framework in this batch with an explicit seed integration.

  4. parallel-code → openspec: parallel-code also has an openspec/ directory — same pattern.

  5. swarmclaw → openclaw/clawhub: SwarmClaw is part of a broader OpenClaw/ClawHub ecosystem, distinct from the SwarmClaw npm package itself. The swarmclaw-skill.md enables OpenClaw agents to control SwarmClaw.

  6. dorothy → skills.sh: Dorothy installs skills from the skills.sh ecosystem, indicating integration with a separate third-party skill registry.

  7. sword-swarm: References a GitHub repo SWORDIntel/claude-backups in the README quickstart — the original repo may have been renamed to SWORDOps/SWORDSwarm. Low credibility signals (24 stars, no license, default branch maine).

← Batch 7
Batch 9 →