Why we built Pulse — and didn't adopt OpenClaw or Hermes

The honest version, because that's the only version worth telling.

The landscape converged — and that's validation, not a threat

By early 2026 agentic orchestration had crystallized around a recognizable shape: a central gateway that multiplexes sessions, dispatches tool calls, sandboxes execution, and drives a multi-turn reasoning loop. OpenClaw (100k+ stars, viral out of late 2025) and Hermes (Nous Research, 140k+ stars, the most-used agent on OpenRouter) are the two largest expressions of it.

Pulse — our dispatch/orchestration engine — is the same shape: a gateway, a governed tool surface, sandboxed headless execution, capability-named dispatch, release gates. We converged on that architecture independently. When the two most popular projects in the field arrive at the design you'd already built, that's your instincts confirmed — not a reason to abandon ship.

Why we didn't adopt OpenClaw

Two reasons — one principled, one technical.

1. Sovereignty. OpenClaw was acquired by OpenAI in February 2026. Our entire thesis is a vendor-neutral, multi-model, self-hosted stack — we migrated across Azure, GCP, and our own hardware precisely to prove nothing is captured. Putting our orchestration substrate under a single model vendor's control is the exact coupling we've spent two years engineering away from. Principled non-starter.
2. Risk posture. OpenClaw's strength is breadth — shell, filesystem, browser, Docker, and a wide array of messaging surfaces, all reachable by the agent. That's maximal reach and a correspondingly large attack surface (the subject of its own published security analyses). Pulse goes the other way: capability-named dispatch, gVisor-sandboxed execution, and SecureVault-mediated credential grants so agents get scoped access, not raw shell. Governed least-privilege over maximal convenience — a deliberate choice for a system that operates real businesses and real money.

Why not just use Hermes, then?

Fair question, honest answer. Hermes is genuinely aligned with our values: open, self-hosted, model-agnostic (Claude / GPT / Gemini / Qwen / DeepSeek), with persistent cross-session memory. If we were choosing a runtime substrate from scratch today, Hermes would be a serious candidate.

But Hermes, OpenClaw, and Pulse all solve the same layer: the agent runtime — dispatch, tool-use, memory. That layer is becoming commodity. Where Netrun is actually differentiated is the layer above it — the governed organization: role-scoped C-suite agents with real mandates and escalation authority; the SDLC the agents are bound by; anti-fabrication and output-fidelity rules; a decision register with owners, rationale, and expiries; a multi-phase board-meeting protocol. None of the runtime frameworks provide that — it's the org, not the dispatcher.

And that governance layer is framework-agnostic. It doesn't care whether Pulse, Hermes, or something else runs underneath. We're not religious about the substrate — we'd happily run the boardroom on Hermes if a deployment called for it. The dispatcher is replaceable; the organizational discipline is the invention.

In one line

The runtime is converging toward commodity — and we're glad, because it confirms the instinct and locates our real value elsewhere. We keep the substrate sovereign and vendor-neutral on principle (which rules out OpenClaw-under-OpenAI and makes open, model-agnostic Hermes genuinely interesting). What we built that the frameworks don't is the governed company that runs on top — and it runs anywhere.

Netrun Systems. Framework facts: OpenClaw — open-source agent runtime, ~100k★, acquired by OpenAI Feb 2026. Hermes Agent — Nous Research, open/self-hosted, model-agnostic, persistent memory, ~140k★. Pulse — Netrun's swarm-native dispatch engine (gVisor sandbox, capability dispatch, MCP tool surface, cloud-agnostic).