Conductor CEO Charlie Holtz Walks Us Through His AI Coding Setup

Key Concepts

• Conductor: An orchestration app for managing multiple AI coding agents.
• Agentic Workflow: A paradigm where the human acts as a "conductor" of an orchestra of AI agents rather than writing code manually.
• Workspaces: Isolated environments for experimentation, functioning as abstractions around Git worktrees.
• Token Maxing: A strategy of maximizing AI model usage (often in "fast mode") to accelerate development.
• Malleable Software: The vision that software should be moddable by users, similar to video game mods.
• Caveman Mode: A manual override mode allowing direct file editing within the app.
• Slot-Free Zones: Specific areas of a codebase reserved for human-only interaction to prevent AI-driven "code rot."

---

1. Operational Methodology and Workflow

Charlie, co-founder of Conductor, emphasizes a workflow where the human is the orchestrator.

• Task Initiation: Users trigger tasks via command-line shortcuts (e.g., Command+N), speaking natural language requests to agents.
• Parallel Processing: The app allows users to manage multiple workspaces simultaneously. While one agent works on a PR, the user can switch to another workspace to review code or provide feedback.
• The "Conductor" Metaphor: The user acts as a CEO/Conductor, monitoring agent progress via a dashboard, providing corrections, and merging finished work.
• Enforced Workflow: Conductor mandates a strict process: agents create a worktree, generate a PR, and the user merges it. Direct file editing is discouraged, though "Caveman Mode" exists for rare manual interventions.

2. Technical Infrastructure and Tools

• Tech Stack: Built using Tauri (Rust backend, native Safari web renderer) with a 90-95% TypeScript codebase. The web app is built on Elixir/Phoenix.
• Hardware: Charlie uses a high-spec machine (128GB RAM) to run local models like Parakeet for text-to-speech, though he experiments with low-spec hardware to ensure software efficiency.
• Integrations: Uses Telegram for remote agent communication and MCP (Model Context Protocol) for documentation retrieval.
• Agent Selection:
  • Claude (Opus): Used for creative tasks and back-and-forth collaboration.
  • Codex: Used as the "workhorse" for heavy lifting, debugging, and high-volume tool calls.

3. Engineering Philosophy and Best Practices

• Skills Files & Cloud MD: The team maintains extensive "skills files" (hundreds of lines) that define engineering practices, preventing the AI from defaulting to generic "enterprise" code patterns.
• Human-AI Boundaries: Charlie argues against letting AI act as the primary architect. He advocates for "slot-free zones"—sections of the codebase where AI is forbidden from contributing to ensure human oversight and prevent "vicious cycles" of AI-generated bad code.
• UI/UX Conviction: The team avoids A/B testing and analytics, relying on "gut feel" to ensure the interface feels "crafted." They prioritize spatial, visual interfaces over terminal-only workflows, noting that humans are inherently visual creatures.

4. Data and Research Insights

• Token Expenditure: During the early stages of Conductor (July 2025), Charlie reported a peak spend of $22,000 on tokens in a single month, resulting in tens of thousands of lines of code.
• Code as "Sawdust": A core argument presented is that code is becoming a byproduct ("sawdust") of the prompting process. The real value lies in the prompts and the architecture, not the specific lines of code, which can be regenerated as models improve.

5. Notable Quotes

• "You should feel like the CEO of a little company and you can see all your agents working for you."
• "Code is almost like sawdust now... you’re putting time into describing what you want and how you want it to be built."
• "If you let the AI make your UI choices for you, you can end up with something that just doesn't feel like crafted."

6. Synthesis and Conclusion

The Conductor approach represents a shift from "coding" to "orchestrating." By enforcing a structured workflow—where agents operate in isolated workspaces and humans maintain high-level control—the team aims to solve the chaos of managing multiple AI agents. The future of software, according to Charlie, lies in malleable, moddable systems where the human defines the intent and the AI handles the implementation, effectively treating the codebase as a dynamic, replaceable asset rather than a static artifact.