Batch 24 Notes
Batch Theme: Director/Coder Split + Plan-Mode + Cross-Runtime Workflow Orchestrators
Roster
| # | slug | display_name | stars | distribution_type | tier | key_differentiator |
|---|---|---|---|---|---|---|
| 1 | compound-engineering | Compound Engineering | 17,199 | cross-runtime plugin (6 tools) | A | 38 skills + 50+ persona-md agents; parallel fan-out code review (10-20 agents); ce-compound memory loop; cross-runtime: Claude Code/Codex/Cursor/Copilot/Droid/Qwen; highest star count in batch |
| 2 | flow-next | flow-next | 615 | claude-code plugin + bundled CLI (flowctl) | A | Spec-first (.flow/specs/fn-N-slug.md); source-tagged capture [user]/[paraphrase]/[inferred]; bundled flowctl binary (not globally installed); fresh-context worker per task; decay-aware compounding memory; Ralph autonomous mode |
| 3 | open-multi-agent | open-multi-agent | 6,251 | npm package (@open-multi-agent/core) | A | Goal-to-DAG coordinator; code-class agent definitions; 10 built-in LLM providers; post-run HTML dashboard; one-shot execution model |
| 4 | pluqqy | pluqqy | 21 | Go binary + TUI | A | Context composition manager (NOT orchestration); pluqqy CLI with Bubble Tea TUI; pipelines compose components → PLUQQY.md; zero agents/hooks/orchestration; dormant |
| 5 | barkain-workflow-orch | barkain-workflow-orchestrator | 60 | claude-code plugin | A | Hook-enforced delegation (adaptive nudge escalation); native plan mode as orchestration handoff boundary; DONE|{path} return protocol; Python/Ruff PostToolUse validators; CLAUDE_MAX_CONCURRENT=8 |
| 6 | meta-agent-teams | meta-agent-teams | 2 | shell + markdown | A | 230 pre-built teams; meta-agent/auditor/constitution triad; max 30% change per cycle; git-backed team evolution; 1 portable prompt that works cross-tool |
| 7 | spec-ops-agents-file | spec-ops-agents-file | 0 | methodology-doc | A | Pure methodology: AGENTS.md + AGENTS-slim.md; spec-first for legacy modernization; stop-when-ambiguous rule; GitHub Action for spec change tracking; zero primitives |
| 8 | devteam-michael-harris | DevTeam: Claude Code Multi-Agent Dev System | 11 | claude-code plugin | A | LLM-type prompt hook (unique in batch); Bug Council (5 parallel analysts); complexity-scored model escalation (0-14 scale, haiku→sonnet→opus); SQLite per-task model history; 127 agents across 24 categories |
| 9 | enterprise-ai-agents-spec | AI Agents in Enterprise Product Development Spec | 51 | methodology-doc + roo-code plugin generator | A | Fluid Dependency Graph with typed edges; Delivery/Dependency Agent as distinct role; committee orchestration (Paxos/Raft/PBFT); dual-audience spec (human + AI-parseable); AI-generated dynamic dashboards |
| 10 | squad-framework | squad-framework | unknown | unknown | C | No public repository found; Reddit/Instagram reference only |
Intra-Batch Patterns
Director/Coder split implementations are highly varied. The batch contains four distinct mechanisms: (1) Roo Code switch_mode() delegation (enterprise-ai-agents-spec), (2) Claude Code native plan mode as handoff boundary (barkain-workflow-orch), (3) Sprint Orchestrator → Task Loop → specialists hierarchy (devteam-michael-harris), and (4) goal-to-DAG coordinator spawning specialist agents (open-multi-agent). These represent fundamentally different architectural choices for the same conceptual split.
Hook governance divides the batch cleanly. Claude Code-specific plugins (barkain, devteam, flow-next) all use hooks; cross-runtime frameworks (compound-engineering, open-multi-agent) and methodology documents (enterprise-ai-agents-spec, spec-ops-agents-file) do not. DevTeam is the only framework in the batch with a "type": "prompt" (LLM-evaluated) hook — a pattern distinct from the standard "type": "command" shell hooks used by all other hook-using frameworks.
Model routing is a differentiator. Only two frameworks implement explicit multi-model routing: devteam-michael-harris (0-14 complexity score across 5 factors, haiku/sonnet/opus tiers, escalation chains per tier) and compound-engineering (cross-reviewer role spanning different models). Open-multi-agent supports 10 providers but uses the same model for all agents. The rest are single-model or model-agnostic.
Spec-first has two archetypes. flow-next enforces spec-before-code as a workflow gate with source-tagged capture and spec file naming conventions. spec-ops-agents-file treats the AGENTS.md file as the spec artifact with a stop-when-ambiguous rule. Both achieve spec-first outcomes but through completely different mechanisms.
Stars are a weak quality signal in this batch. compound-engineering (17,199 stars) and open-multi-agent (6,251 stars) dominate by popularity but represent different use cases (cross-runtime plugin vs. npm library). Meta-agent-teams (2 stars) has architecturally novel ideas (git-backed team evolution, 230 pre-built teams) that its star count completely obscures.
Pluqqy is a deliberate non-orchestrator. Despite batch theme inclusion, pluqqy is a context composition manager, not an agent orchestrator. Its inclusion confirms the theme spans adjacent tooling. Its dormant status and minimal stars suggest it did not find product-market fit.
Most Interesting Finds
devteam-michael-harris — The "type": "prompt" hook entry — using Claude to semantically evaluate whether a file modification is in scope before allowing it — is architecturally novel in the entire corpus. All other hook-using frameworks use shell commands for hook actions. The Bug Council pattern (5-analyst parallel escalation for stuck bugs) is also unique. The 0-14 complexity scoring algorithm with per-task model history in SQLite is the most explicit cost-optimization strategy documented in any single framework.
enterprise-ai-agents-spec — The Fluid Dependency Graph with formally typed dependency edges (Definition, Implementation, Testing, Production) is the most formally specified data structure in this batch. The separation of the Delivery/Dependency Agent (owns FDG + dashboards) from the Orchestrator (owns task scheduling + policy enforcement) is a clean architectural pattern absent from other frameworks. The dual-audience spec design (human-readable + AI-parseable with formal math notation) is a unique authoring choice that positions the spec itself as agent instruction.
barkain-workflow-orch — The adaptive nudge escalation (stderr reminders that intensify before becoming blocking) as the primary governance mechanism is the most pragmatic approach in the batch. Using native Claude Code plan mode as the planning/execution boundary eliminates the need for a separate planning command or PRD artifact.
Tier C Items
| slug | reason |
|---|---|
| squad-framework | No public repository found despite exhaustive search across GitHub; referenced in Reddit/Instagram only |
Cross-Batch Connections
- devteam-michael-harris is the closest to claude-flow (Batch 5) in component count and complexity but is a Claude Code plugin rather than an npm library, and uniquely adds model escalation on failure.
- enterprise-ai-agents-spec shares the Roo Code target with frameworks from Batches 3 and 16 but is the only framework that specifies committee-based orchestration with named consensus protocols.
- flow-next shares spec-first enforcement and fresh-context workers with openspec-derived frameworks (Batch 1) but implements these through a bundled CLI and decay-aware memory rather than skill files.
- meta-agent-teams shares the constitutional constraints pattern with spec-ops-agents-file (both have immutable rules governing agent behavior) but implements it through a 3-agent triad (meta-agent + auditor + constitution) rather than a static document.