Behind the session

The Agent Stack

Who built this, what model did the thinking, and where does the inference actually run?

Hermes Agent v0.19.0

A personal AI agent by Nous Research that runs the same agent core across a CLI, a messaging gateway (Telegram, Discord, Slack, and ~20 other platforms), a TUI, and an Electron desktop app. It learns across sessions (memory + skills), delegates to subagents, runs scheduled jobs, and drives a real terminal and browser.

In this session, Hermes ran inside the desktop app on Jon's Windows 11 machine (RTX 5090, 32GB VRAM). It used its terminal, file, browser, vision, and code execution tools to drive tldraw's local HTTP API and ComfyUI's API — all local, all on one machine.

Version: v0.19.0 Upstream: commit 2b0fb72a Install: D:\AI_software\Hermes Session: 2026-07-28

GLM-5.2 — The Language Model

The LLM that powered every decision in this session — reading Jon's instructions, writing ComfyUI prompts, generating Python scripts, composing HTML, and analyzing images with vision. GLM-5.2 is Zhipu AI's (operating globally as Z.ai) latest flagship model, built specifically for coding, reasoning, and tool-driven agentic workloads.

Parameters: 744B (Mixture-of-Experts) Context: 1,000,000 tokens Output: 131,072 tokens License: MIT (open-weight) Effort: High
  • Agent-oriented design: Built for autonomous, multi-step tool use — not just chat. First-class function calling and multi-turn tool use.
  • IndexShare architecture: Reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at 1M context length.
  • Speculative decoding: Improved MTP (Multi-Token Prediction) layer increases acceptance length by up to 20%.
  • Anti-hack RL training: Specifically trained to resist reward hacking in coding agent trajectories.
  • Open weights: MIT-licensed, available on Hugging Face — no regional limits, no technical access borders.

engy.ai on the Bittensor Network

GLM-5.2 inference was served by engy.ai, a provider running on the Bittensor network — a decentralized AI infrastructure protocol. Instead of a single cloud provider (AWS, Azure, Google), Bittensor distributes AI compute across a network of independent operators who compete to provide intelligence commodities.

How Bittensor works:

  • Subnets: The network is organized into specialized subnets, each focused on one type of AI task (text generation, image generation, storage, etc.). Each subnet is a self-contained, incentive-driven marketplace.
  • Miners (41% of emissions): Run inference — they execute the actual AI model computations and serve responses to clients.
  • Validators (41% of emissions): Score the miners' work for quality, ensuring the network rewards good inference and penalizes bad.
  • Subnet creators (18% of emissions): Define the subnet's task and incentive rules.
  • TAO token: The bridge between clients and compute. TAO is minted every block and distributed to miners, validators, and subnet creators based on performance.
  • dTAO (dynamic TAO): Introduced February 2025 — each subnet gets its own "Alpha" token and liquidity pool. The more TAO flowing into a subnet, the higher its Alpha value, and the larger its share of daily emissions. Market-driven allocation.
  • Post-halving (Dec 2025): Daily TAO issuance cut from 7,200 to 3,600, tightening supply while the subnet economy expands.
  • Fixed supply: 21 million TAO total — same cap as Bitcoin.

Bittensor is, at its core, a horizontally scalable, cryptography-based resource-allocation protocol for intelligence commodities. Subnets like Chutes (serverless inference backed by real GPU compute) and Targon are projecting millions in annual revenue, demonstrating that decentralized AI infrastructure can compete with centralized cloud providers at a fraction of the cost.

When Hermes Agent made an LLM call in this session, the request went to engy.ai, which routed it through a Bittensor subnet to a miner running GLM-5.2 inference on their GPU hardware. The miner served the response, the validator scored it, and TAO was emitted as reward — all happening in seconds, invisibly, behind every tool call and every line of generated text.

The Full Stack

Jon Tojek (VFX Artist, XYZ Graphics)
↓ speaks to
Hermes Agent v0.19.0 (desktop app, Windows 11, RTX 5090)
↓ powered by
GLM-5.2 (744B params, 1M context, MIT license, effort: High)
↓ served by
engy.ai → Bittensor network (decentralized inference, TAO token)
↓ drives
tldraw offline v1.12.0 (localhost:7236) + ComfyUI (localhost:8188)
↓ generates with
KREA2 turbo fp8 (local, RTX 5090, zero API spend)