The Ecosystem

Module 0.2 · The Landscape · Master Course

15 minutes · Prerequisite: Module 0.1 (three-job definition; thickness spectrum)

From "a harness" to "the field"

Module 0.1 gave you the invariant: loop + tools + safety, varied by thickness.

But "harness" still spans an 80-line personal assistant and a 368,000-star platform with 40 messaging channels deployed by NVIDIA and Microsoft. Reading both with the same mental machinery is a category error.

The fix is a taxonomy — not to be exhaustively correct, but to be predictive.

The seven categories

CategoryWhat it tells you
Terminalloop in a process you start; tools are fs/shell
Platformloop behind a gateway; channel-aware trust
Meta-harnessinherits a base; contributes routing + subagent hierarchy
Orchestration frameworkno "the loop" — a state machine you compose
Hardened forkinherits everything; adds governance OUTSIDE the agent
Academicoptimized for inspectability, not production
SDK / runtimeYOU are the harness author; it gives you primitives

Each category predicts where the interesting engineering lives → which rubric module to read first.

Read the field through lineage, not feature lists

A feature list tells you what a harness does. A lineage map tells you what it disagrees with. The second is more honest.

Aider (2023) ── git-first, Polyglot Benchmark
  └ influenced most later CLI harnesses' git integration
Pi (2024) ── ~1.2k LOC, 4 tools
  └ oh-my-pi
OpenClaw (2025) ── 368k stars, 40+ channels
  ├ NemoClaw (NVIDIA, Apr 2026) ── governance fork
  ├ Microsoft Scout (Jun 2026) ── Entra identity fork
  └ Hermes (Feb 2026) ── COMPETITOR, not fork
      └ `hermes claw migrate` (migration FROM OpenClaw)
OpenCode (2025)
  └ oh-my-opencode (Sisyphus + Prometheus + Atlas)
OpenAI Agents SDK ── sibling to Codex CLI (Rust)

The diff between parent and child reveals design intent. Studying a fork without its parent loses the meaning.

OpenClaw vs Hermes: breadth vs depth

OpenClaw — breadth

40+ messaging channels
NVIDIA + Microsoft partnerships
#1 by total usage until May 2026

Architectural debate: cross-channel trust boundaries

→ teaches Module 6 + Course 2 trust model

Hermes — depth

Layered persistent memory
Self-evolving skills
Overtook OpenClaw on OpenRouter May 10 2026

Architectural debate: memory poisoning surface

→ teaches Module 4 memory architecture

Both are correct. They optimized for different buyers. Ranking one above the other mistakes a design decision for a quality score.

The governance move: NemoClaw & Scout

The most important idea in this module for where the course is going.

  ┌─ Agent process (REACHABLE by the agent) ──────┐
  │  Loop   Tools   Permission flags / sysprompt  │
  └────────────────────┬──────────────────────────┘
                       │  every call passes through
  ┌────────────────────▼──────────────────────────┐
  │ GOVERNANCE LAYER  (OUTSIDE the agent's reach) │
  │  NeMo Guardrails · OpenShell · Entra identity │
  └────────────────────┬──────────────────────────┘
                       │
                       ▼
                 World (fs, net, channels)
If the agent can reach the enforcement layer, a compromised agent can disable it. NemoClaw and Scout put enforcement outside the agent's reach. This is the foundation of Course 2's entire threat model.

Meta-harness, as a routing workflow

Per the course visual stack: n8n first. "Meta-harness" = classify → dispatch → aggregate.

       Task Intake ("pentest X" / "review PR Y" / "solve CTF Z")
            │
            ▼
      Classify Task   ◀── meta-harness job #1 (the router)
            │
       ┌────┼────┐
       ▼    ▼    ▼
     recon review ctf   ◀── inherited sub-harnesses (NOT replaced)
       │    │    │
       └────┼────┘
            ▼
   Aggregate + Cost Ledger  ◀── meta-harness job #3 (one evidence log)
            │
            ▼
         Report

Same pattern in oh-my-opencode (Sisyphus→Prometheus→Atlas), the CSI meta-harness, and Capstone F. One workflow, three courses.

Study priority: where to spend your time

🔴 ESSENTIAL deep-dive exhaustively (the course's 21 deep-dives)

🟠 HIGH sample each for its one distinct idea

🟡 MONITOR be aware; full study optional

Priority is not quality. A 🟡 harness may be excellent — it is 🟡 because it is not architecturally distinctive enough to warrant a dedicated session, not because it is inferior.

Three anti-patterns

1. Ranking by stars. Stars measure attention, not architecture. OpenClaw's 368K tells you nothing about whether its trust architecture is sound (it isn't — that's why NemoClaw exists).
2. Treating categories as quality tiers. "Terminal" is not lesser than "platform." Pi (terminal) is one of the most important harnesses in the field.
3. Studying forks without the parent. You cannot understand NemoClaw without OpenClaw. A fork's meaning is its diff from the parent.

Takeaways

  • Seven categories, each predictive of where the interesting engineering lives.
  • Read the field through lineage, not feature lists or star counts.
  • OpenClaw vs Hermes = breadth vs depth. Both correct.
  • NemoClaw & Scout: governance beneath the agent. Foundation of Course 2.
  • Study priority (🔴🟠🟡) decides where your time goes.

Next: Module 0.3 — the rubric and methodology you'll apply to every harness in this ecosystem.

Harness Engineering Master Course · Module 0.2 · 03-slide-deck