Architecture Patterns

Thin Harness + Thick Skills

The harness provides minimal core infrastructure (context management, memory, message routing), while all domain logic lives in modular, installable skills.

┌── Harness (thin) ──────────────┐
│  • Message routing              │
│  • Context window management    │
│  • Memory persistence           │
│  • Safety / permission layer    │
│  • Skill discovery & loading    │
└────────────┬───────────────────┘
             │
    ┌────────┼────────┐
    ▼        ▼        ▼
 [Skill A] [Skill B] [Skill C]
  GitHub    Calendar   Deploy
  Ops       Mgmt       Pipeline

Pros: Modular, community-extensible, model-agnostic Cons: Skill quality varies, coordination complexity Examples: OpenClaw, Nexu

Monolithic Harness

All agent logic is built into a single, tightly-integrated system.

Pros: Deeply optimized, consistent behavior, easier to debug Cons: Vendor lock-in, hard to extend, model-locked Examples: Claude Code, Cursor Agent

Harness-as-a-Service

The harness runs in the cloud, managed by a platform. Users configure but don't host.

Pros: Zero ops, always available, managed scaling Cons: Data leaves your machine, platform dependency Examples: Claude Managed Agent, Codex cloud

Three Architectures Visualized

graph LR
    subgraph "Thin Harness"
        TH[Harness Core] --> S1[Skill A]
        TH --> S2[Skill B]
        TH --> S3[Skill C]
        TH --> M1[Any Model]
    end

    subgraph "Monolithic"
        MH[Full Harness<br/>All Logic Built-in] --> M2[Single Model]
    end

    subgraph "HaaS"
        Cloud[Cloud Platform] --> MH2[Managed Harness]
        MH2 --> M3[Platform Model]
    end

Comparison: Claude Code vs Codex vs OpenClaw

Dimension	Claude Code	Codex	OpenClaw
Harness size	~512K lines	Unknown (closed)	~50K lines
Model support	Claude only	GPT only	Any model
Memory	Platform-managed	Encrypted summaries	User-owned files
Skills	Built-in tools	Built-in tools	Community skills
Customization	CLAUDE.md	AGENTS.md (limited)	AGENTS.md + MEMORY.md + Skills
Sandbox	Docker-based	Cloud sandbox	Local + Docker
Open source	No (source visible)	No	Yes (MIT)
Multi-agent	Limited	Yes (Codex tasks)	Yes (sub-agents)

Which Pattern Fits?

If you need...	Choose...
Maximum control & customization	Thin harness (OpenClaw/Nexu)
Best single-model experience	Monolithic (Claude Code)
Zero setup, cloud-first	HaaS (Managed Agent)
Team of agents working together	Thin harness with multi-agent support

---

Next: Memory Systems →