Conductor Founders on Pioneering Multi-Agent Coding

Founders' Path from College to YC Startup

Charlie and Jackson met in college, where Charlie (a fifth-year student) taught Jackson weightlifting and tried recruiting him for CS classes. They reconnected post-graduation—Jackson at Netflix on machine learning infrastructure, Charlie at Replicate handling growth and engineering. Bonding over side projects and ultimate frisbee (a 'serious team' they likened to an elite sports dynamic), they decided to start a company after an interview with Aaron Epstein. Initially applying to YC with an AI reservation-booking idea (e.g., using Sonnet 3.5 for browser control like reserving tennis courts), they quickly pivoted, recognizing it as 'a solution in search of a problem.' Jackson recalls lacking a clear startup vision but craving a change from remote work; Charlie had long dreamed of ditching bosses after reading Hacker News mantras like 'you weren't meant to have a boss.' Their partnership thrives on complementary styles: Charlie as the rapid 'hacker' prototyping duct-taped MVPs, Jackson as the 'gardener' refining for robustness.

YC Iteration: 10+ Ideas to Self-Solve Dev Friction

During YC, they spent the first 1.5 months 'casting around,' building prototypes friends hated and switching ideas every few days—impressing Aaron with build speed (shipping new concepts faster than others' MVPs). Pivotal office hours advice from Aaron: focus on dev tools and 'make something these guys want,' leading them to poster their faces with that mantra. Obsessed with AI coding potential post-Sonnet 3.5, they eyed tools like Aider (grandfather of Claude/Codex) for post-IDE workflows where humans direct AI to execute tasks. Early attempts failed—models required too much handholding, tool calling was nascent. They pivoted to Chorus, a multi-model chat app solving their own needs, releasing it in January post-batch. Building Chorus's dev tools uncovered Conductor: manually juggling 5 repo clones with Claude/Codex hit friction, evolving to worktrees for isolation.

"Put up a poster on your wall of your guys' faces and then just say across it at the top 'make something these guys want' and just build something that you guys want." — Aaron Epstein's advice, which Charlie and Jackson literally followed.

Prototyping Conductor took 2-3 weeks: Charlie solo-built MVP in a week, Jackson iterated despite initial UI critiques ('this is not at all what I wanted'). Validation hit when running parallel agents yielded productive results—'unread dot' notifications signaled completion, and using Conductor to build itself was 'exhilarating,' ripping through independent features/bug fixes.

Conductor's Core: Parallel Agents and Cloud Unlock

Conductor is a Mac app consolidating coding agents (Claude, Codex) outside IDE/terminal: one-click codebase isolation via worktrees, task assignment, review/merge. Pioneering multi-agent runs (beyond 1-2), it grew 10x since January, attracting indie hackers to Big Tech engineers. Launching Conductor Cloud today enables persistent workspaces—agents continue post-laptop shutdown, addressing instance limits (users maxing 3-5). Charlie notes mental limits on managing agents; future interfaces need higher abstractions. Early manual processes (pre-app) proved multi-agent viability once models hit a 'sweet spot' needing partial oversight.

"The first time giving the agent a task then pressing command N and giving another agent a task... seeing a little unread dot come up... it had finished. That was pretty magical." — Jackson on the 'aha' of parallel agents.

Insights from Top Engineers and Viral Growth

Bike tours to user sessions reveal elite setups: heavy investment in 'skills files' (Markdown with codebase-specific best practices, e.g., React tips), surprisingly vanilla configs (avoiding Vim-like over-customization memes), and 'slot-free zones'—AI runs wild in isolated areas, humans architect core. Advice evolves rapidly; current YC batch sees different agent strategies monthly. Charlie's Replicate growth role honed viral demos, like GPT-4 Vision's David Attenborough-narrated webcam hack (viral despite Attenborough's dislike). Tips: authenticity (raw screens, direct language like talking to friends), reps (post often), replicate what you heart on Twitter—'make something you want to click on.'

"The best engineers are putting a lot of thought into their skills files... simple and vanilla setups... slot-free zones where you let the AI run wild in certain parts." — Charlie on observations from frontier engineers.

Key Takeaways

• Build for yourself first: Solve personal pain (e.g., multi-agent friction) before scaling; use mantras like 'make something these guys want.'
• Iterate fast in YC: Ship prototypes every 3 days, pivot after 10 users try (don't theorize).
• Complement skills: Hacker (rapid MVPs) + gardener (robust polish) dynamic accelerates product-market fit.
• Multi-agent coding: Use worktrees for isolation, skills files for context, cloud for persistence; cap at 3-5 mentally, abstract interfaces next.
• Top engineer habits: Vanilla setups, targeted skills docs, human-as-architect in core zones.
• Viral demos: Raw/authentic (no corporate speak), observe/emulate what excites you on social, high reps.
• Business momentum: 10x growth from indie to enterprise; $22M Series A fuels team expansion.

"I think we have a nice dynamic where there's like two archetypes of engineers... myself as a hacker... Jackson as a gardener." — Charlie on their partnership.

"Observe what you like... when you scroll and try and replicate that... make something that you want to consume." — Charlie/Jackson on viral content strategy.