Engineering at AI Speed

Lessons from the First Agentically Accelerated Software Project

QCon San Francisco 2025 Conference Session

---

About This Repository

This repository contains the presentation materials and live demos for the talk "Engineering at AI Speed: Lessons from the First Agentically Accelerated Software Project."

The talk draws from our experience building Claude Code, Anthropic's CLI for AI-assisted development—the first AI coding agent built using itself. 90% of Claude Code's code is written with or by Claude Code, giving us unique insights into what it means to work at AI speed.

The Message

Implementation is becoming free. Feedback loops are becoming everything.

The shift isn't that AI writes code faster—it's that you can now optimize for speed of learning rather than quality of judgment.

Three Stories

The talk explores three real implementations from Claude Code:

1. Rebuilding Input 🎯 - 333 → 945 lines over 9 months. Custom text input with Unicode handling. Success through testable architecture.

2. Reimagining Shell 🐚 - Pivoted from persistent to transient shells. Discovered snapshot pattern through experimentation.

3. Reversing SQLite 🗄️ - 15 days from launch to removal. Learned architectural problems couldn't be predicted from debate.

Running the Presentation

This presentation uses a custom terminal-based presentation system with live code demos.

bun install
bun src/presentation.ts

Navigation

• Arrow Keys (←/→): Navigate between slides
• Arrow Keys (↑/↓): Select interactive links
• Enter: Activate links (run demos, open URLs)

Live Demos

The presentation includes interactive demos that run in tmux:

• demos/cursor-*.ts - Text input and Unicode handling
• demos/shell-*.ts - Command execution and environment management
• demos/sqlite-*.ts - Database complexity and migration issues
• demos/simple-storage.ts - The simple JSONL replacement

Repository Contents

content/
├── src/               # Presentation system implementation
├── static/            # Presentation content and ASCII art
│   ├── presentation.md        # Main slide deck
│   ├── presentation-notes.md  # Speaker notes
│   └── *.txt                  # ASCII art assets
├── demos/             # Live code demonstrations
└── docs/              # Technical documentation
    └── markdown.md    # Presentation system docs

The Framework We (Me & 🤖) Built

The presentation itself runs on a custom terminal-based presentation system we built for this talk. It features:

• Markdown-based slide format
• ASCII art rendering
• Live code execution via tmux integration
• Interactive navigation
• asciicinema playback

Three Questions to Ask

When evaluating any experiment:

1. What does shipping reveal that planning can not? Domain constraints, user workflows, second-order effects

2. What architectures enable evolution? Testability, composability, minimal dependencies

3. What distinguishes detours from dead ends? Does pain decrease or increase with each iteration?

The Practical Playbook

1. Ship small, ship often - Target hours not days
2. Make reversal cheap - Feature flags, modular architecture
3. You have to unship - Don't rationalize wrong turns, reverse them

Contact

Adam Wolff Member of Technical Staff, Anthropic

• Twitter/X: @wolffiex
• GitHub: wolffiex