newsfilter.io
Product Demonstration, Interview

Conductor CEO Charlie Holtz Walks Us Through His AI Coding Setup

Y CombinatorCharlie Holtz
Watch on YouTube →

Conductor Overview & Current State

  • Charlie, co-founder of Conductor (YC Summer 24), developed an app to orchestrate multiple coding agents on macOS.
  • Conductor currently builds itself, with 90–95% of the desktop app written in TypeScript despite a Rust backend.
  • The web application is built on Elixir (Phoenix) and is currently minimal, serving only as a login interface.
  • The core desktop architecture utilizes the native Safari web renderer for the UI layer.

Workflow & Operational Decisions

  • Users "kick off" new tasks via Command-N, speaking directly to the AI to request task execution (e.g., reviewing Linear issues).
  • Workflows enforce a strict non-terminal interface: workspaces function as work trees, creating Pull Requests (PRs) that must be reviewed and merged within Conductor.
  • Direct file editing is restricted; the app defaults to an AI-driven process where humans only intervene via specific "Caveman Mode" or highlighting text for AI modification.
  • Status tracking is segmented into three states: in progress, in review (once a PR is created), and done (after merging).
  • A "Cloud Workspaces" feature is under development to enable agents to run in cloud environments rather than being constrained by the local Mac's CPU.
  • Conductor implements "slot-free zones," requiring human review for specific lines of code to prevent AI from entering a cycle of generating bad code from bad code.
  • The team maintains a "dangerously accept all permissions" default for Claude to ensure agents can operate without friction.
  • The interface design prioritizes human visual cognition over terminal efficiency, organizing chats, editing, and code review into a spatial layout.
  • High-level architectural and UI decisions are made exclusively by humans to prevent AI from generating un-crafted software structures.

Agent Selection & Model Strategy

  • Claude (Opus) is preferred for creative feature building and back-and-forth collaboration.
  • Codex is designated as the "workhorse" for debugging, handling large volumes of tool calls, and solving specific technical problems.
  • The team employs "Fast Mode" by default to maximize token efficiency and encourages high-effort prompting strategies.
  • Context windows are expanded using a Context 7 MCP server to fetch necessary documentation during tasks.
  • Custom "skills" files (e.g., cloud.md) define engineering practices and boundaries, often hundreds of lines long, to guide agent behavior.

Usage Statistics & Cost

  • In July 2025, the team spent $22,000 on tokens using previous model generations, representing a peak in spending during the initial launch phase.
  • The team actively avoids "token maxing" (spending on lines of code) and prioritizes generating minimal, high-quality code over volume.
  • Local models, specifically Parakeet for text-to-speech, run on a machine equipped with 128GB of RAM.
  • The founder recently purchased a base-model MacBook Neo to enforce reliance on low-spec hardware and external workflows.

Future Roadmap & Philosophy

  • The "dashboard" feature aims to provide a CEO-level view of all active agents, allowing for digestible reporting and direction setting.
  • Future iterations will move toward "malleable software," allowing users to mod workflows and agents similar to video game mods.
  • The team is experimenting with multi-player chat capabilities where multiple humans collaborate simultaneously with AI agents.
  • A "Gary Mode" has been introduced to visualize all tool calls by default for a specific power user.
  • The philosophy posits that code is becoming "sawdust" while prompts and descriptions become the primary asset for software generation.
  • The team intends to keep core infrastructure on human-written APIs and contracts, leaving only specific, non-critical chunks open for AI experimentation.
  • Development is guided by "gut feel" and daily internal usage rather than traditional A/B testing or analytics.

Community & External Integration

  • External users have reverse-engineered Conductor to create mobile versions by spoofing Inter-Process Communication (IPC) calls.
  • Telegram integration is actively used for specific open-source interactions.
  • Spokenly is used locally for text-to-speech with a focus on privacy and offline capabilities.
  • The "gooseneck microphone" is cited as an essential hardware tool for whispering commands to maintain office quietness in open-plan settings.