Marc Partensky · AI infrastructure

AI infrastructure: the harness, the models, the tools, the privilege

One operator, one harness, several model providers, thirty-five tool servers, and agents that act on real systems instead of chatting about them.

Surveyed 15 November 2026 · public page: internal addresses, ports and credentials are deliberately omitted

1 · the harness one process that holds the context, the jobs and the surfaces hermes-agent 2026.9.24 3 profiles: default, discord-bridge, pocket memory and skills 31 skills loaded on demand, persistent memory, session store scheduling cron jobs, webhooks, queues, delegation to sub-agents surfaces Matrix, Discord, mobile, web UI, desktop app
2 · the models routed per task, with a fallback that is measured, not assumed deepseek default · 1M context · vision cheap enough to run all day openrouter fallback, per task routes when the first provider is down kimi-k3 auxiliary vision reads screenshots and scans claude opus + sonnet heavy reasoning, code delegated, not in the hot path 2 LAN endpoints on the laptop and the handheld local models, no API cost
3 · the tools thirty-five MCP servers, two transports, one shared process infra nixos, git, artifacts object storage, torrents devices on the LAN self-hosted apps files and calendar mail, short links flashcards, media money bank accounts, payments, market data personal and OSINT identity checks, web investigation, public datasets Two transports: local stdio for servers that only speak stdio, HTTP on the private network for the rest. One process hosts the whole tool plane: 8 servers and 90 tools in 85 MB, against 133 processes and 1.7 GB with one process per session.
4 · the part that makes it usable an agent that can act is only useful if acting is bounded privilege is brokered no sudo, no ambient credentials: a request, a proof, a record changes are gated rebuilds serialized, impure evaluation refused, rollback is a store path same repo as the host the agent edits the config it runs on, and cannot merge without a human

Why this is worth a look

Most agent projects stop at the prompt and the demo. The hard part is the tool plane: how an agent reaches thirty-five real services without holding thirty-five credentials.

The second hard part is privilege. Here root is a service with two possible proofs, a scope, a rate limit and an audit trail, so an autonomous agent can act on a production-like host without being trusted with it.

The third is cost. Provider routing, a shared tool process and a spend dashboard per model and per key are what make an agent affordable to run every day instead of in a test.

Sources: the harness config (profiles, model routing, 35 tool servers), its session store, the running services on the host, and a tailnet survey. Verified 15 November 2026.
Catppuccin Mocha · Lavender accent