New to this topic? That is completely fine. It is completely fine if you have never studied some of these topics before. We want to see how fast you can learn something new, grasp it, and teach it from first principles.
This page is a quick primer, not a script. Use it to get your bearings, then read the links below and anything else you find: other blogs, papers, videos, code. We are judging the story you build, your examples and metaphors, your first-principles thinking, and any experiments you run yourself.
A quick primer
A language model on its own can only do one thing: take text in and predict text out. It cannot read your files, run code, remember yesterday or stop itself from doing something dangerous. The harness is all the software around the model that turns it into an agent that can actually get work done.
The agent loop
Every harness, from Claude Code to Codex to Pi to Hermes, runs some version of the same loop:
The interesting questions are in the design choices. What goes into the system prompt? Which tools, and how many? What happens when the conversation gets too long (compaction)? Does the agent remember things across sessions? Who is allowed to approve a dangerous command?
Two very different answers: Pi and Hermes
| Pi | Hermes Agent | |
|---|---|---|
| Philosophy | A minimal, extensible harness you make your own. Deliberately leaves out features like sub-agents and plan mode. | A “self-improving” agent with a built-in learning loop. |
| Built by | Mario Zechner. TypeScript, MIT licence. | Nous Research. Python, MIT licence. |
| Core tools | Four by default: read, write, edit, bash. | 40+ tools in toolsets: web, terminal, browser, media, memory, scheduling. |
| Memory | Branching session files with compaction, plus instruction files like AGENTS.md. | Persistent memory, cross-session search, and a model of the user. |
| Skills | Loads skills on demand. | Writes new skills from experience and improves them over time. |
| Where it runs | Terminal UI, RPC and an SDK. | Terminal plus a gateway to Telegram, Discord, Slack, WhatsApp, Signal and email, with a scheduler. |
| Safety | No built-in permission system; the docs recommend a sandbox. | Command approval and container isolation. |
A quick orientation only, from each project's README and docs on 9 November 2026. Both projects move fast, so go to the repos and docs below for the real details.
The repositories
- Pi (earendil-works/pi) · Mario ZechnerThe Pi monorepo: unified LLM API, agent loop, coding-agent CLI and terminal UI.
- Pi documentation · pi.devHow the loop, sessions, compaction, extensions and skills work.
- Hermes Agent (NousResearch/hermes-agent) · Nous ResearchThe self-improving agent: memory, skills, messaging gateway, scheduler.
- Hermes Agent documentation · Nous ResearchTools and toolsets, skills, memory, MCP and architecture.
Read
- What I learned building an opinionated and minimal coding agent · Mario Zechner (creator of Pi)Why Pi is built the way it is. Note: Pi has added features since this post, so check the current docs.
- The Anatomy of an Agent Harness · LangChainEverything outside the model is the harness.
- Agent Harness Engineering · Addy OsmaniHarness engineering as a discipline.
- Building effective agents · AnthropicWorkflows versus agents, and the core patterns.
- Effective harnesses for long-running agents · AnthropicWhat a harness does beyond the model when work spans many sessions.
Watch
- ▶ Building pi in a World of Slop · Mario Zechner, AI Engineer
- ▶ I Hated Every Coding Agent, So I Built My Own · Mario Zechner, Mastra
- ▶ Hermes Agent Fundamentals in 29 Minutes · Tina Huang
- ▶ Nous Research's Hermes Agent: The Case for Open Models in Production · Arize Observe 2026
Go beyond this list. These are starting points. The best submissions usually bring in something we did not link: an experiment, a paper, a story of their own.