Solving Real Friction: Building Conductor for Themselves

Charlie and Jackson, Conductor's co-founders, started as college teammates who reconnected post-graduation—Jackson at Netflix on ML infra, Charlie at Replicate on growth and engineering. They quit stable jobs to build together, inspired by ultimate frisbee's elite team vibe: "putting your all into it and working against big challengers." Their dynamic shines: Charlie as the rapid "hacker" prototyping duct-taped MVPs, Jackson as the "gardener" refining into robust systems.

In YC, they cycled through 12 ideas, starting with AI booking reservations (e.g., tennis courts via Sonnet 3.5 browser control)—a "solution in search of a problem." Prototypes bombed; friends hated them. Aaron Epstein's advice clicked: poster with their faces saying "Make something these guys want." Pivoting to dev tools, they obsessed over Aider (pre-Claude Code precursor) and Sonnet 3.5's overhang. Early Conductor attempts failed—models needed too much handholding, tool calling was nascent.

They shipped Chorus, a multi-model chat app, in January post-batch. Building it exposed pain: manual multi-agent workflows with repo clones and worktrees. Charlie prototyped Conductor in days: one-click isolated codebases, parallel Claude/Codex runs, review/merge UI. Jackson's litmus test: "This is not at all what I wanted... but this is really impressive." Three weeks from first code to release, they ripped through features dogfooding it—"building Conductor with Conductor" felt magical.

"The first time giving the agent a task then pressing command N and giving another agent a task... seeing a little unread dot... it had finished," Charlie said. That parallel productivity hooked them; users from indie hackers to Big Tech engineers followed, growing 10x since January.

Unlocking Scale: Conductor Cloud and Interface Challenges

Conductor's core: Mac app consolidating agents beyond IDE/terminal limits. Create isolated worktrees, dispatch tasks, monitor via unread dots, merge surgically. Pre-launch, users maxed local instances (3-5 manageable mentally). Today's announcement: Conductor Cloud—agents persist post-laptop shutdown.

Bottlenecks persist: humans cap at 3-5 agents without better interfaces. "We've proven you can run more than one... but to get past three to five, we'll need a higher level abstraction," Charlie noted. Early hacks: five repo clones running Claude. Worktrees streamlined it. Future: slot-free zones (AI wild in sandboxes, humans architect core).

Business traction: $22M Series A (Spark/Matrix), small team scaling fast. They bike-tour SF, watching frontier engineers.

Viral Demos and Authentic Growth Hacks

Charlie's Replicate growth role honed demos: GPT-4 Vision launch hack—webcam + David Attenborough voice clone narrating like Planet Earth—went viral (even Attenborough saw it). Tips: raw screen recordings, direct language ("talking to a friend"), no corporate speak. Reps matter: post tons, observe what you heart on X, replicate emotions.

"Make something that you want to click on," Jackson echoed YC wisdom. Authenticity scales: show Python screens unpolished, prove what's possible.

Top Engineers' Agent Playbooks: Simple Setups, Deep Skills

Watching elite users: setups are "surprisingly simple and vanilla"—no Vim-meme complexity. Key differentiator: skills files. Markdown docs codify codebase lore (e.g., React best practices from vets). Agents ingest them per task.

"The best engineers are putting a lot of thought into their skills files," Charlie observed. Hybrids emerge: AI autonomous in periphery (slot-free zones), humans on architecture. Even YC advice evolves monthly amid model leaps.

"You can easily meme here and put a ton of thought into configurations... but not necessarily get work done."

Key Takeaways

  • Build for your own friction first: Charlie/Jackson solved multi-agent pain manually before coding Conductor—led to PMF in weeks.
  • Ship prototypes every 3 days in YC: Iterate fast, even if ideas flop; speed separates founders.
  • Test litmus: Does your co-founder use it? Jackson's buy-in validated amid shiny distractions.
  • Viral content formula: Raw demos + direct voice + reps; make what you'd click.
  • Top engineers prioritize skills files over configs: Feed agents React patterns, codebase specifics via Markdown.
  • Scale agents via interfaces: Conductor Cloud persists work; aim for 3-5+ with abstractions/slot-free zones.
  • Team dynamics win: Hacker (rapid MVPs) + gardener (robust polish) = elite output.
  • Dogfood ruthlessly: Ripping code with your tool confirms magic.
  • Pivot on action, not theory: Talk fails; build prototypes, watch drop-off.