Orchestrating Code: A Look Inside Conductor's Development Workflow

Charlie, co-founder of Conductor, an application that allows users to orchestrate multiple coding agents on their Mac, shares insights into his daily workflow and the philosophy behind building their product. Having gone through the Y Combinator Summer 2024 batch, Charlie demonstrates how Conductor is used to build itself, emphasizing a shift from traditional coding to a more directive, AI-assisted approach.

The Evolution of Interaction: Talking to Computers

The modern workspace, especially with open-plan offices, can be distracting. Charlie highlights a simple yet impactful addition to his setup: a $20 gooseneck microphone. This tool encourages more direct verbal interaction with computers, making it less disruptive to lean over and whisper commands like "Please merge PR 3475" to an AI assistant. This move towards voice interaction is a key trend, aiming to make communicating with technology more natural and efficient.

Building Conductor with Conductor

Charlie spends a significant portion of his day within Conductor, using the application to develop Conductor itself. His workflow involves frequently initiating new tasks by pressing Command + N. This action, a sneak peek into Conductor's upcoming cloud workspaces feature, allows him to verbally prompt the AI: "Can you take a look at the latest Linear issue and give me a rough pass at how you'd solve it?"

Once a task is initiated, it appears in the sidebar, allowing Charlie to multitask. He utilizes keyboard shortcuts extensively, such as Command + Shift + Y to quickly access a workspace ready for merging. While he reviews the code generated by Claude, he often provides feedback. If Claude doesn't get things exactly right, Charlie leaves comments, similar to GitHub's style, asking clarifying questions like, "This looks a little bit weird to me. Why do we need this?" He then lets Claude run with the feedback and moves on to another workspace.

Experimentation is a core part of Charlie's process. He constantly kicks off new workspaces to explore different ideas, acknowledging that most don't make it into the final product. This iterative approach allows for rapid prototyping and testing of concepts, with successful experiments potentially being promoted through internal and experimental settings.

Conductor On the Go

Conductor is also extending its capabilities beyond the desktop. Charlie demonstrates the "on the go" feature by speaking into his phone: "Let's add a new feature where I can change the theme to hacker mode." After clicking "Conduct," his computer begins working on the request, enabling him to manage tasks remotely.

The Diminishing Role of Manual Coding

When asked if he still writes code manually, Charlie admits it's rare. His manual coding is typically limited to minor edits, such as adjusting Tailwind classes or modifying environment files. Conductor even includes a "caveman mode" for situations where direct keyboard input is necessary, but this is explicitly named to signify its infrequent use. For most small edits, he prefers to highlight code and provide verbal instructions to the AI or simply speak commands like, "That button looks a little too wide, can you make it smaller?"

The application's interface is designed for seamless integration. A workspace ready for merging can be archived with a single click, and completed tasks are moved into a "done" folder. A new "status" feature in the left panel tracks tasks through their lifecycle: "in progress," "in review" (once a PR is created), and "done" (after merging).

A new dashboard page aims to provide a centralized view of all agent activities, allowing users to direct their agents effectively. The ideal user experience is to feel like the CEO of a small company, overseeing agents who provide digestible reports and await direction for corrections or approval.

Essential Tools and Customizations

Charlie's daily software stack includes Telegram for communication with Claude and Spokenly for text-to-speech, running a local Parakeet model. His powerful computer, with 128GB of RAM, facilitates running local models. He also mentions ordering a low-spec MacBook to challenge himself to use minimal resources.

Key customizations that matter include meticulously crafted "skills files" and cloud.md documents. These contain hundreds of lines of instructions, including specific engineering practices tailored for a startup environment, differentiating their approach from enterprise coding.

He consistently uses "fast mode" for token efficiency and the "context 7 MCP" for documentation retrieval. A critical, non-default setting is "dangerously accept all permissions" when running Claude, which is central to Conductor's operational philosophy.

"Slop-Free" Zones and AI Boundaries

A core principle at Conductor is maintaining "slop-free zones" – areas of the codebase or documentation that are known to be human-written. While AI can contribute, every line must be human-reviewed. This prevents AI from entering a negative feedback loop, where it generates bad code based on previous AI-generated errors. Some codebases even include explicit "do not touch if you are an AI" directives.

Conductor's Tech Stack

Conductor is a Tauri app, utilizing the native Safari web renderer. The backend is written in Rust, but the desktop application is predominantly TypeScript (90-95%). The web app is a small Phoenix app built with Elixir, currently limited to login functionality, though Charlie expresses a strong preference for Elixir.

Guiding the AI: Don't Let It Be Your Architect

Charlie emphasizes that users should not let AI dictate architectural decisions. Concepts like Conductor's "workspace" feature, while an abstraction, were human-conceived. Similarly, UI/UX decisions, such as the layout of chats and sidebars, were carefully considered to create a crafted and intuitive experience. Allowing AI to make these design choices can lead to a less polished product.

A potential area for improvement is building the core of the app around human-written APIs and contracts, with distinct areas for AI experimentation. This separation ensures that AI exploration doesn't compromise core infrastructure. Conductor aims to push user comfort zones, a strategy that initially met with skepticism regarding managing multiple AI agents. The enforced workflow, where workspaces must create PRs for merging, was a deliberate choice to guide user interaction.

Adapting to Evolving Models

The challenge and excitement lie in adapting to the rapid advancements in AI models. As agents become more intelligent and capable of running for longer periods, the underlying infrastructure needs to evolve beyond the constraints of a user's local machine.

Building Conviction in Opinionated Design

Conductor's opinionated design approach is built on conviction, primarily through daily use by the development team. If something doesn't feel right, it's quickly addressed. This gut-feel approach, rather than relying heavily on analytics or A/B testing, ensures the product feels intuitive and cohesive.

Codex vs. Claude: Choosing the Right Tool

While Conductor supports both Codex and Claude, Charlie finds himself using Codex more frequently for its workhorse capabilities in tackling specific problems and performing extensive tool calls. Claude, on the other hand, is preferred for more creative, back-and-forth interactions, acting as a more collaborative partner, especially when building new features. Codex is the go-to for getting tasks done efficiently.

Beyond the Terminal: The GUI Advantage

The move from terminal interfaces to GUIs in the 1980s highlights humans' spatial and visual nature. Charlie argues that while AI might thrive in a command-line environment, humans benefit from visual cues. The ability to see chats on the left, the main interaction in the center, and a review panel on the right provides a more intuitive and less restrictive experience than a pure terminal interface. GUIs also unlock functionalities not possible in a terminal.

Token Maxing and Code Minimalism

Charlie shares a high token spend of $22,000 in July 2025 during Conductor's early development, using previous-generation models. He advocates for "token maxing" and using "fast mode" for high-effort interactions. However, Conductor aims to keep lines of code minimal to prevent codebase sprawl. The approach to code volume differs significantly between starting a new app and working on an established codebase like Conductor.

Workflow Evolution: Six Months of Change

Compared to six months prior, Charlie now relies less on opening an IDE for manual changes on complex PRs. He also uses the GitHub web app significantly less, as code review and commenting can be done directly within Conductor. The addition of a "checks tab" allows for integrating GitHub checks directly into Conductor.

Surprising Uses of Conductor

The most surprising use of Conductor was a user building a mobile version by spoofing inter-process communication calls to the desktop app. Gary, a power user, has also significantly pushed the boundaries of Conductor's capabilities, particularly with its "skills" feature. Conductor even has a "Gary mode" that keeps tool calls expanded for better visibility.

The Future of Collaboration and Code

Charlie believes the future lies in enhanced collaboration between humans and AI, including multiplayer chats where multiple users work with AI on the same task. The metaphor of conducting an orchestra resonates: guiding AI agents at a high level, intervening for specific adjustments, and allowing them to perform in unison.

Code, he suggests, is becoming akin to "sawdust" – a byproduct of the process rather than the primary focus. The emphasis is shifting from crafting code to describing desired outcomes. This means prompts become paramount, and with new model generations, rerunning prompts can yield updated code.

The concept of "malleable software," similar to modding video games, is key. Users should be able to customize Conductor with their own workflows, while the core structure remains polished and well-thought-out. This allows users to make the software feel truly their own.