Phase D Batch 17 — Multi-agent cowork/desktop platforms

Roster (10)

Intra-batch Patterns

UI Metaphor Divergence is the sharpest axis of variation across this batch. Every framework solves the same problem (surfacing multi-agent activity to humans) but with a fundamentally different UI metaphor:

1. Terminal TUI (agent-deck): Bubble Tea Go TUI; no browser required; keyboard-driven; closest to developer tooling
2. Electron kanban (aperant): Drag-and-drop board for 12 parallel Claude Code worktrees; visual task tracking
3. Electron custom chat (eigent, open-claude-cowork, hola-os): Electron wrapping a custom React chat/workspace UI
4. Next.js web portal (tinyagi, multica): Browser-based multi-agent management; tinyagi's TinyOffice at :3000, multica's full web app
5. Matrix IM client (hiclaw): Element Web at :18088; agents and humans share a real IM room — the UI is not built by the framework
6. Task management board (paperclip): React+Vite board at :3100; deliberately not a chat interface; org chart, issues, budgets
7. Full-stack platform (lobehub): Next.js web + Electron desktop + CLI; closest to a SaaS platform; IM gateway to Discord/Telegram/etc.

Orchestration Architecture clusters into three types:

• Manager-Workers (hiclaw): explicit Manager role; all routing through Manager; Matrix room as bus
• Platform-managed dispatch (lobehub): no fixed Manager; platform assembles agents for each task; schedule drives execution
• Org-chart hierarchy (paperclip): corporate reporting structure; heartbeat engine drives execution per agent
• Parallel orchestration (aperant, agent-deck, eigent, multica): multiple agents run simultaneously on partitioned tasks; results merged
• Swarm / peer-to-peer (tinyagi): [@agent: message] direct peer addressing; no coordinator required

Credential / Security Architecture shows interesting variation:

• Higress AI Gateway (hiclaw): only framework with a dedicated API proxy hiding real credentials from agents — enterprise-grade credential isolation
• Standard env vars (all others): agents receive API keys directly

Persistence Layer varies widely:

• SQLite (agent-deck, open-claude-cowork)
• PostgreSQL + pgvector (multica)
• Embedded PostgreSQL (paperclip)
• Full PostgreSQL + Drizzle (lobehub)
• MinIO + Matrix room history (hiclaw)
• SQLite workspace state (hola-os)
• None / ephemeral (eigent)

Most Interesting Finds

1. hiclaw's Matrix protocol bet: Using Tuwunel (self-hosted Matrix server) + Element Web as the agent communication bus is architecturally unique in the entire corpus. The orchestration bus is a real federated IM protocol — any Matrix client can be used as the UI, agent-to-agent and human-to-agent communication use the same protocol, and room history is the audit log. Combined with Higress AI gateway credential isolation and Kubernetes CRDs, hiclaw is the most enterprise-architecturally distinctive framework analyzed in Phase D.

2. paperclip's anti-chat UI: The explicit product decision to use a task management board (not a chat window) as the primary interface is a rare example of a framework taking a strong anti-pattern stance and building architecture around it. The "agents have jobs, not conversations" thesis is internally consistent from the org chart model to the heartbeat engine to the board UI.

3. lobehub's self-iteration tool: @lobechat/builtin-tool-self-iteration enables agents to modify their own system prompts based on observed feedback. Combined with white-box personal memory, this is the only framework in the batch with an explicit agent self-modification primitive. The "co-evolution" framing (agents and humans improving each other) is backed by actual architectural components, not just marketing copy.

4. tinyagi's swarm protocol: The [@agent: message] peer-to-peer message addressing pattern — where any agent can send a direct message to any other agent without routing through a coordinator — is the only decentralized (non-hierarchical, non-platform-managed) multi-agent communication pattern in the batch.

5. aperant's QA loop with exit codes: Aperant's QA agent (12th parallel worker) validates all other 11 agents' outputs and can reject them, blocking completion until quality criteria are met. This is the only framework in the batch with a dedicated validation agent as a structural component rather than an optional check.

Items Written as Tier C

None. All 10 frameworks had sufficient source code and documentation for full 11-file analysis. Note: open-claude-cowork (@anthropic-ai/claude-agent-sdk@0.2.6 patched) is the simplest framework in the batch — effectively a 1-file Electron wrapper over the Claude Agent SDK — but was still fully analyzed.

Cross-References Discovered

• hiclaw ↔ higress: hiclaw is from the agentscope-ai organization, which is part of the Higress/Alibaba open-source ecosystem. The Higress AI Gateway is hiclaw's central infrastructure component.
• lobehub ↔ lobechat: LobeHub is the rebranded/extended version of LobeChat (the AI chat app with 60k+ stars). The architecture inherited LobeChat's provider abstraction layer and UI component library.
• paperclip ↔ openclaw: Paperclip explicitly positions itself relative to OpenClaw ("If OpenClaw is an employee, Paperclip is the company"). The adapter-openclaw-gateway is a first-class supported adapter.
• eigent ↔ camel-ai: Eigent is a desktop UI for CAMEL-AI's Python multi-agent framework. The Python backend (@camel-ai/camel) provides the actual agent runtime; Eigent adds the Electron GUI layer.
• tinyagi ↔ discord/telegram: TinyAGI's swarm communication channels include Discord, Telegram, and WhatsApp as first-class agent-to-human and agent-to-agent channels alongside the Matrix-like room model.
• multica: Most vendor-neutral framework in the batch — explicitly supports 11 different agent CLIs (Claude Code, Codex, Gemini, Cursor, Aider, Devin, Jules, Copilot, OpenClaw, Continue, Tabnine) as workers in its Squad routing pattern.