Agentic Architecture in Practice — Harness · LLM · Directives

A reveal.js slide deck presenting a working model for how agentic systems are designed, governed, and extended, using RockBot (rockbot.dev) as the running example.

By Rockford Lhotka · blog.lhotka.net

▶ Play the deck live: https://marimerllc.github.io/rockbot-presentation/

(Arrow keys to navigate, F for fullscreen, S for speaker notes, Esc for the slide overview. Published via GitHub Pages — every push to main rebuilds it.)

Abstract

An agent is not just an LLM with tools bolted on. In practice, it's the combination of three things: a harness that controls execution, an LLM that reasons and generates, and directives that shape behavior. This talk uses RockBot as a concrete example of that architecture and shows how the pieces fit together in a real system. From there, we'll look at what becomes possible once the core is solid: memory that preserves context, skills that package repeatable behavior, self-learning and skill evolution that improve the system over time, and integration layers like MCP and A2A that let agents reach tools and collaborate with other agents. The point is to move from abstract "agent" talk to a working model for how agentic systems are actually designed, governed, and extended.

The demo throughout is RockBot — rockbot.dev.

The thesis

The deck argues that an agent is not "an LLM with tools bolted on." In practice it's three distinct parts:

Agent = Harness + LLM + Directives

• Harness — the code that controls execution: the loop.
• LLM — reasons and generates. A black box.
• Directives — shape behavior: scope, tone, rules.

Pull any one out and you don't have an agent. From there the deck shows how getting the core right lets the system extend: memory, skills, self-learning, scheduled tasks, "dreams," MCP, A2A, and least-privilege governance.

Outline

1. Title — Agentic Architecture in Practice
2. What is an agent? (the equation)
3. The LLM is a black box
4. The harness is the loop
5. Directives shape behavior
6. Meet RockBot
7. Tools: how the harness acts
8. Peers: agents calling agents (A2A)
9. The harness handles a lot
10. Memory: three tiers + a graph
11. Skills: fixed guides & evolving know-how
12. Tasks on a schedule
13. LLM routing: tiers & fallback
14. The orchestrator delegates down (primary · subagents · workers · wisps)
15. RockBot dreams
16. Principle of least privilege
17. Untrusted code runs sandboxed
18. The working model (recap)

Viewing the deck

Open index.html in a browser. The deck loads reveal.js 5.1.0 from a CDN, so an internet connection is needed for the slide framework; local theme.css and assets/ supply the custom styling and imagery.

Because the deck is fetched over file:// you can usually just double-click index.html. If a browser blocks local asset loading, serve the folder over HTTP:

npx serve .
# or
python -m http.server

Exporting to PDF

A pre-rendered export is included as rockbot-agentic-architecture.pdf.

To regenerate it:

npm install          # installs Playwright + Chromium
npm run pdf          # runs export-pdf.mjs

export-pdf.mjs drives headless Chromium against reveal.js's ?print-pdf mode to produce the PDF.

Files

Path	Purpose
index.html	The reveal.js deck
theme.css	Custom theme / styling
assets/	RockBot logo and favicon
export-pdf.mjs	Playwright PDF export script
rockbot-agentic-architecture.pdf	Pre-rendered PDF of the deck
package.json	Scripts and dev dependencies