# AgentClaw Architecture Comparison Assessment

**Source:** https://agentclaw.space/blog/13-ai-employees-architecture  
**Date:** 2026-04-21  
**Status:** Analysis complete — adoption recommendations below

---

## Executive Summary

AgentClaw's 13-agent system shares core principles with HoffDesk but differs significantly in scale, communication patterns, and memory architecture. This assessment identifies what we can adopt versus what we already do better.

---

## Side-by-Side Comparison

| Aspect | AgentClaw (13 Agents) | HoffDesk (3 Agents) | Assessment |
|--------|----------------------|---------------------|------------|
| **Agent Count** | 13 specialized | 3 generalist/specialist hybrid | AgentClaw wins on scale |
| **Identity System** | `employees/AGENT.md` per agent | `SOUL.md` + `IDENTITY.md` | **Parity** — both markdown-based |
| **Memory Architecture** | `employees/memory/AGENT/YYYY-MM-DD.md` | `memory/YYYY-MM-DD.md` | **Ours is simpler** — one shared timeline |
| **Shared State** | `shared-brain/*.json` (21 files) | `shared/` directory | **AgentClaw wins** — structured JSON handoffs |
| **Communication** | File-based async only | Telegram Boardroom + `shared/` | **Ours wins** — synchronous when needed |
| **Spawn Protocol** | `spawn-employee.sh` injects identity+memory | OpenClaw native spawning | **Parity** — similar outcome |
| **Vector Search** | Yes (knowledge base) | No (memory search only) | **AgentClaw wins** — semantic search across all history |
| **Databases** | 8 SQLite + vector | Minimal (blog SQLite only) | AgentClaw heavier, we leaner |
| **Self-Improvement** | Explicit lesson logging | Implicit via MEMORY.md | **Parity** — similar intent |
| **Orchestration** | Fully automated handoffs | Human-in-the-loop (@mentions) | **Different philosophies** |

---

## What AgentClaw Does Better

### 1. Structured JSON Handoffs (`shared-brain/`)

**Their approach:**
```json
// intel-feed.json
{
  "discovered_at": "2026-03-04T14:30:00Z",
  "trending_topic": "local AI deployment",
  "source": "twitter",
  "priority": 1,
  "assigned_to": "SCRIBE"
}
```

**Our approach:** Markdown prose in `shared/project-docs/`

**Adoption recommendation:** ✅ **YES — Create structured handoff JSONs**

Create `shared/handoffs/` for machine-readable agent communication:
- `content-queue.json` — pending articles with metadata
- `design-requests.json` — Daedalus tasks with specs
- `sprint-status.json` — active work items across agents

### 2. Vector Knowledge Base

**Their approach:** Semantic search across all agent history, lessons, outputs

**Our approach:** `memory_search` only (recent sessions + MEMORY.md)

**Adoption recommendation:** ⚠️ **PARTIAL — Investigate for future**

Our current scale (3 agents, 1 user) doesn't warrant the complexity. At 6+ agents, vector search becomes essential. **Defer until scaling decision.**

### 3. Automated Trend Detection

**Their approach:** TRENDY agent scouts content every 2 hours, auto-populates intel

**Our approach:** Manual content ideas in `content-ideas.md`

**Adoption recommendation:** ⚠️ **OPTIONAL — Consider for content pipeline**

Could add a "scout" agent for Home Lab / OpenClaw communities. Not critical path.

### 4. Explicit Lesson Logging Protocol

**Their approach:** Mandatory "lessons learned" write after every task

**Our approach:** Ad-hoc MEMORY.md updates

**Adoption recommendation:** ✅ **YES — Formalize post-task reflection**

Add to AGENTS.md: "After significant work, append lessons to memory/YYYY-MM-DD.md"

---

## What HoffDesk Does Better

### 1. Human-in-the-Loop Orchestration

**Our approach:** Telegram @mentions for coordination, Matt approves major decisions

**Their approach:** Fully automated handoffs

**Assessment:** ✅ **KEEP OUR APPROACH**

Faisal's system is optimized for content mills (viral posts, tweets). Our system prioritizes quality, security, and sovereignty. Human judgment at key decision points is a feature, not a bug.

### 2. Zero-Trust Security Model

**Our approach:** Tailscale-only, no external exposure, local-first architecture

**Their approach:** Cloud-hosted 24/7 agents (AgentClaw is a cloud service)

**Assessment:** ✅ **SIGNIFICANT ADVANTAGE**

Our Sovereign Constraint is a deliberate architectural choice AgentClaw cannot easily replicate.

### 3. Clear Agent Boundaries (Socrates/Daedalus/Wadsworth)

**Our approach:** Strict separation of concerns via SOUL.md

**Their approach:** 13 agents with overlapping responsibilities

**Assessment:** ✅ **MORE SUSTAINABLE**

Three well-defined agents scale better than 13 thinly-defined ones. Our routing is cleaner.

### 4. Simpler Memory Architecture

**Our approach:** One `memory/` directory, shared timeline

**Their approach:** Per-agent memory directories + shared brain

**Assessment:** ✅ **EASIER TO MAINTAIN**

At 3 agents, unified memory is simpler. At 13, their approach becomes necessary.

---

## Recommended Adoptions

### Immediate (This Week)

| Change | Implementation |
|--------|---------------|
| **Formalize lesson logging** | Update AGENTS.md with post-task reflection protocol |

### Near-term (Next Month)

| Change | Implementation |
|--------|---------------|
| **Content pipeline tracking** | Enhance `content-ideas.md` with clearer status workflow |

### Deferred (Until 6+ Agents)

| Change | Trigger |
|--------|---------|
| **Structured JSON handoffs** | When markdown prose becomes coordination bottleneck |
| **Vector knowledge base** | When `memory_search` becomes insufficient |
| **Per-agent memory isolation** | When cross-agent memory pollution becomes problematic |
| **Automated trend detection** | When content volume justifies scout agent |

**Note on JSON handoffs:** Currently unnecessary. Markdown prose in `shared/` works for 3 agents + human coordination. JSON state files become valuable when:
- Agent count exceeds 5 (human coordination limit)
- Automation needs to make routing decisions without human read
- Programmatic filtering/sorting of work items becomes essential

**Hybrid structure ready if needed:**
```
shared/
  project-docs/          # Markdown: thinking, specs, research
  handoffs/              # JSON: state (deferred until scale)
```

---

## Hierarchy Comparison

### AgentClaw Command Structure
```
Founder (Faisal) → Spawn Script → 13 Agents → Shared Brain
                     ↑___________________________↓
                         (automated feedback loop)
```

**Philosophy:** Minimal founder involvement, maximum automation

### HoffDesk Command Structure
```
Director (Matt) → Wadsworth → Routes → Socrates / Daedalus
                     ↑________↓
                (human approval gates)
```

**Philosophy:** Founder as strategic decision-maker, agents as execution

---

## Key Insight

> "The constraint is no longer execution capacity. It is judgment — knowing what to build, what to cut, and what to direct the agents toward next."

**AgentClaw and HoffDesk agree on the fundamental premise.** We differ on implementation:

- **AgentClaw optimizes for volume** (content production, viral marketing)
- **HoffDesk optimizes for quality + sovereignty** (infrastructure, correctness, security)

Neither is wrong. AgentClaw's architecture is appropriate for their use case. Ours is appropriate for ours.

---

## Action Items

| Task | Owner | Priority |
|------|-------|----------|
| Create `shared/handoffs/` directory structure | Wadsworth | P1 |
| Draft `content-queue.json` schema | Wadsworth | P1 |
| Update AGENTS.md with lesson logging protocol | Wadsworth | P2 |
| Create `shared/brain-index.md` registry | Wadsworth | P2 |
| Evaluate vector DB options for future | Socrates | P3 (defer) |

---

## Files

- Assessment: `/shared/project-docs/research/agentclaw-comparison-assessment.md`
- Original source: https://agentclaw.space/blog/13-ai-employees-architecture