Agentic AI with Claude Code and Open Claw

James Rowland
Cape Fear Makers Guild
April 7, 2026

LLMs and the Pretraining Miracle

"It's just a stochastic parrot" — not quite.

  • Predict the next token on trillions of words from the internet
  • No labels needed — the text is the training signal
  • No labeling bottleneck → scale to massive datasets cheaply

Pretraining Input

The quick brown fox jumps over the lazy

Compression = Learning

A model with N parameters trained on ≫ N tokens can't memorize — it must compress

  • Forced to find patterns, structure, and abstractions
  • Regularization through sheer data pressure
  • This is why scaling works: more parameters + a lot more data → deeper understanding

A stochastic parrot memorizes. A model that compresses generalizes.

Compression = Learning

The ChatGPT Moment

Pretrained models know a lot — but they're unpredictable

  • Raw models ramble, contradict, hallucinate, and ignore instructions
  • Brilliant but unruly — like a genius with no social skills

RLHF (Reinforcement Learning from Human Feedback) changed everything

  • Humans rate outputs → model learns what "helpful" looks like
  • Aligns the raw capability to follow instructions and be useful
  • This is what turned a research artifact into a product billions use

RLHF Input

[USER] What is the capital of France? [/USER] [ASSISTANT] The capital of France is Paris. [/ASSISTANT]

Thinking Models

The next breakthrough: let the model reason before answering

  • Model gets a hidden scratchpad — a "thinking block"
  • Reinforcement is based on the final answer, not the thinking
  • The model discovers its own reasoning strategies

Why this matters:

  • No one teaches it how to think — only what's correct
  • It learns to plan, backtrack, and self-correct on its own
  • Unlocks performance on math, code, and complex multi-step problems

Thinking Input

[USER] What is 37×53? [/USER] [THINKING] Let me work this out... 37×50=1850, 37×3=111, so 1850+111=1961 [/THINKING] [ASSISTANT] 1961 [/ASSISTANT]

Tool Use

Models can't interact with the real world on their own — tools fix that

How it actually works (no magic):

  1. Model outputs "I want to call search('weather NYC')"
  2. We intercept that, run the tool, return the result
  3. Model continues with the new information
  • Try things, get feedback, adjust — just like humans
  • Trained during thinking blocks so the model learns when and how to reach for tools
  • Read files, run code, search the web, call APIs — anything you wire up

Tool Use

center

Tool Use Input

[USER] What's the weather in NYC? [/USER] [THINKING] I need the weather [TOOL_USE] search("weather NYC") [/TOOL_USE] [TOOL_RESULT] 72°F, sunny [/TOOL_RESULT] [/THINKING] [ASSISTANT] It's 72°F and sunny in NYC. [/ASSISTANT]

Persistence

Agents are more persistent than humans — they don't get bored or frustrated

The catch: they need to know when to stop

  • The more explicit your exit criteria, the longer an agent can work effectively
  • Vague goals → spinning wheels. Clear "done" → deep iteration.

Context management matters for long-running tasks

  • Models have finite context windows — fill it up and quality degrades
  • Skills: predefined prompts that inject focused instructions without bloating context
  • Subagents: spin up a child agent for a subtask, get back a summary, keep the parent lean

Task Length Is Growing

center

Claude Code Demo

  1. Install and sign in — code.claude.com/docs/en/quickstart
  2. Explore slash commands and skills
  3. Create a pygame game from scratch
  4. Watch it install dependencies automatically
  5. Run the game, hit errors, watch it fix them
  6. Iterate until it works

If It's a CLI, It's a Tool

Agents know how to use the terminal and can help you install and configure CLIs. Agents are text-first, so CLIs make the most sense.

Classic: git, ffmpeg, curl, python, package managers

Modern: gh, discord, gws — interact with the web without a browser

Agents can navigate browsers through image input, but CLIs are faster, more reliable, and token efficient

These Slides Were Built by an Agent

LLMs love Markdown — so these slides were built with Marp, a Markdown-based slide framework, by Claude Code configured with skills for Marp.

The full transcript of the agent session is here:

github.com/rowland-208/marp-decks

Open Claw Demo

  1. Install Open Claw — docs.openclaw.ai/start/getting-started
  2. Set up with OpenRouter
  3. Configure with the web UI

Warning: Open Claw is dangerous by default — it runs with minimal safety guardrails out of the box

CFMB — Our Home Grown Discord Bot

logo

Claude Code and Open Claw are just wrappers around models — no reason you can't build your own