AI Agent Swarms Are Amazing

Joel Hooks · Badass Courses · Day 1

Agent swarms.

Going exponentially with Opus 4.5+.

Harness engineering is the state of the art.

model | harness | context

The Claude Code arch diagram is wild.

Harness includes:

Own the tools.

opencode — gave us the opportunity to build.

Swarm — multi-agent coordination to survive context death.

Anthropic nerfed opencode.

AI Skills — Matt Pocock — mattpocock/skills

OpenClaw!

SOUL.md — give the agent personality, an icon.

Your harness should make you smile.

Pi

Other thing: agent harness called Pi.

Workflow in Pi is better than Claude Code's garbage implementation.

Follow Mario on X.

Slow the F* Down** — don't do everything, and don't do it quickly. (Slow Productivity)

Pi will update itself; it will read its own source code to add stuff. Out of the box is not featureful.

Nico Bailon — Pi stuff — Pi subagents.

pi-cmux — replacement for Terminal/Ghostty, but built on Ghostty. Pi extension to manage it.

Pi — modify yourself to be better.

pi-notes — Thariq — HTML is not better than markdown.

mdsvx — Svelte version of MDX (which is what?).

Human readable, but the agent can read it as well.

wzzrd.sh — notes to share with the team.

Second brain — better notes.

pi-feedback — support inbox, last turn of the session, sends diff to process the feedback for the session.

Key metrics — having a lot of users will lead to a lot of support tickets.

Feedback tool to improve the automated support, lots of turns.

agentdungeon.ai — agents playing DnD, one is the DM, and the other agents work and collaborate.

Cloudflare — Durable Objects — state.

Sunil Pai — follow recommendation. Something called Think.

Dillon Mulroy — artifacts — memory, primitives.

Agent swarms are hard, it's true.

Addy Osmani — The Orchestration Tax.

Cognitive bandwidth is not parallelizable.

The real skill is building the systems that can pull any and all of the threads of the system and context limit.

Slides