How Conductor CEO Charlie Holtz Sets Up His Team Of AI Agents
Charlie Holtz, co-founder of Conductor, demonstrates his AI-agent orchestration workflow where he manages multiple Claude and Codex agents simultaneously on his Mac to build software without writing code himself. He discusses opinionated design decisions, 'slot-free zones' to prevent AI code quality degradation, and his philosophy that code is becoming 'sawdust' — a byproduct of well-crafted prompts. The interview covers token spending, tech stack, and his vision of humans as conductors of AI orchestras.
Summary
Charlie Holtz, co-founder of Conductor (YC Summer 2024), walks through his daily workflow using his own product to build Conductor itself — a Mac app for orchestrating multiple coding agents simultaneously. He uses a gooseneck microphone to whisper commands to Claude in an open-plan office, constantly spawning new agent workspaces via keyboard shortcuts (Command+N) and voice input. He describes a workflow where agents pick up Linear issues, write code, create PRs, and await human review, all visible in a sidebar with statuses (in progress, in review, done).
Holtz rarely writes code himself, describing manual coding as 'caveman mode' — a deliberate feature name in Conductor for the rare occasions a human needs to edit a file directly. He prefers highlighting code and giving verbal feedback, or just describing desired changes conversationally. He spends heavily on tokens, with a peak of $22,000 in a single month (July 2025), and advocates for always using 'fast mode' and maximum effort settings.
A core architectural philosophy is the concept of 'slot-free zones' — sections of the codebase or documentation that are human-written and must be reviewed line-by-line before any AI contribution is accepted. This guards against a vicious cycle where AI sees bad code and generates more bad code. Related to this, Holtz argues humans must remain the architects: AI should not make high-level UI or structural decisions, as those require crafted human judgment.
On tooling, he uses Spokenly (running local Parakeet model) for speech-to-text, the Context7 MCP for documentation, and always runs Claude with 'dangerously accept all permissions' as the default in Conductor. He differentiates Claude Opus (creative, conversational, good for new features) from Codex (a workhorse for grinding through specific problems with many tool calls).
Holtz discusses the broader philosophy behind Conductor's opinionated design: enforced PR-based workflows, no direct file editing, and a 'CEO of a little company' dashboard metaphor where users direct agents and review digestible reports. He sees the product as pushing users slightly past their comfort zone — early feedback called managing even three agents 'crazy.' He also teases future directions including cloud workspaces, longer-running agents, multiplayer human-AI collaboration, and 'malleable software' akin to video game mods, where users can customize and re-prompt software as models improve.
Key Insights
- Holtz argues that code is becoming 'sawdust' — it used to be the crafted structure you were building, but now it's merely a byproduct of well-crafted prompts, meaning when a new model generation arrives you can simply rerun your prompts and get entirely new code, making the old code irrelevant.
- Holtz introduced 'slot-free zones' — deliberately human-written sections of the codebase where every AI-contributed line must be read by a human — because he observed that AI can enter a vicious cycle of generating increasingly bad code when trained on its own prior low-quality output.
- Holtz spent $22,000 on API tokens in a single month (July 2025) while bootstrapping Conductor, and explicitly advocates always running in 'fast mode' with maximum effort settings rather than optimizing for cost.
- Holtz argues that humans must remain the architects of their software and must not let AI make UI or structural decisions, because AI-driven design choices produce interfaces that feel 'uncrafted' — he cites Conductor's three-panel layout as a decision requiring extensive deliberate human thought.
- Holtz differentiates Claude Opus from Codex by use case: he reaches for Opus when building new features because it feels more creative and collaborative, while Codex is the 'workhorse' he uses when the goal is simply grinding through a specific problem with many tool calls.
Topics
Transcript
[0:00] Hello, I am Charlie, the co-founder of Conductor, which is an app that lets you orchestrate a bunch of coding agents on your Mac. And we were YC Summer 24. Uh, and I'd love to show you my setup. So, a recent thing that I can't live without is this uh gooseeneck microphone. $20 on Amazon. we are all trying to talk to our computers more. [0:31] Um, one issue with having like an open floor plan office is that can be pretty distracting. So, one advantage of these is you can like lean over and whisper into Claude and be like, "Please uh merge PR 3475 and it's a little bit less disruptive. I we we all got…
Full transcript available for MurmurCast members
Sign Up to AccessMore from Y Combinator
AI Can't Learn The Way Humans Do - This Could Fix That
The video explains how world models—neural networks that predict future states of an environment given actions—could solve AI's sample efficiency problem and potentially unlock AGI. By learning to simulate environments like humans do, AI systems could train policies on synthetic data rather than requiring millions of real-world examples, as demonstrated in recent robotics and gaming applications.
New Ways To Design With AI Tools
E Bufar, head of design at Y Combinator, demonstrates how she uses AI tools like Claude and Aqua to design products end-to-end, eliminating traditional design workflows. She walks through three projects—Paxel, Sodazine, and Startup School—showcasing techniques like voice-to-code development, comprehensive markdown documentation, and rapid iteration through disposable design tools.
See How Customers Actually Use Your Product
The speaker explains dot plots, a visualization tool that displays individual user behavior over time by plotting user actions on a 2D grid. Unlike aggregate metrics like DAU/MAU, dot plots reveal usage patterns, feature adoption, and retention issues that are invisible in traditional dashboards.
How A Prototype Built During A Missed Flight Became A New Gusto Product
Eddie Kim, Gusto co-founder and head of technology, discusses Gusto Co-founder, an AI product that automates business processes for small business owners. Built in 10 weeks by a 5-person team during a missed flight layover, the product lets customers automate tasks like payroll and scheduling through SMS and Slack, while proactively suggesting optimizations and uncovering hidden opportunities like tax credits.
India Can Create The Largest AI Companies
Y Combinator leaders and Indian venture investors discuss why India is positioned to create the world's largest AI companies, emphasizing that success now depends on technological edge rather than market knowledge, and that AI has democratized startup building for young founders with high agency and customer obsession.