Pithagoras vs Hermes: a local AI agent harness review

Pithagoras is the speaker's agent harness built around Pi for work on local hardware. This walkthrough explains its architecture and compares the speaker's experience with Hermes and OpenClaw. It is a development account rather than a controlled benchmark: the speaker found Hermes consumed inference capacity without making its activity clear enough.

The project began as a web portal that let a coding assistant keep working after the browser closed. It grew to include Telegram, pluggable channels, cron schedules and memory. Understory connects through a Pi MCP adapter; a memory file provides a fallback. Separate work and autonomous sessions keep project coding apart from channel and scheduled tasks. The scheduler has no default heartbeat.

The speaker reports roughly 70 tokens per second with a tuned Qwen 35B model using MTP, falling to about 30 per request with two parallel requests. A 128,000-token context budget informs which tools each session receives. Debugging examples cover accidental new conversations after restarts, a Telegram polling deadlock and identity files that needed explicit explanations of their purpose.

Deployment has clear limits. The speaker intends access through a LAN or Tailscale and says authentication and access control have received little work. Telegram and webhooks have been tested; Slack and Discord have not been tried with real credentials. Browser disconnects leave tasks running, but a portal restart stops the active run while retaining its conversation.