Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
The Tool the Best Engineers Are Using Right Now
TL;DR
Conductor is betting that engineers will manage fleets of coding agents, not just one assistant in an IDE — Charlie and Jackson describe their Mac app as a control layer for Claude Code, Codex, and other agents, and launched Conductor Cloud so jobs keep running even after you shut your laptop.
The company has hit real traction fast — fresh off a $22 million Series A co-led by Spark and Matrix, the team says Conductor has grown roughly 10x since January and is used by everyone from indie hackers to engineers at large public companies.
Their core insight came from using the pain themselves — before Conductor, they were manually juggling five repo clones, then Git worktrees, and discovered the 'magic moment' when one agent could finish a task while they were already assigning work to another.
YC advice directly shaped the product that worked — after burning through many ideas, Aaron Harris pushed them toward devtools and told them to 'make something these guys want,' which led them to build for their own workflow instead of forcing a shaky AI reservation-booking idea.
The best AI engineers aren't necessarily the ones with the fanciest setup — from watching top users, Jackson says the strongest pattern is thoughtful 'skills files' in markdown that encode codebase-specific knowledge, while many elite engineers keep the rest of their tooling surprisingly simple and 'vanilla.'
They think today's bottleneck is human cognition, not just model capability — Charlie says he can only truly manage about three to five agents at once, and argues the next breakthrough is interface design and review workflows, since we're still stuck in a '2010 GitHub PR review era.'
The Breakdown
From local agent orchestration to Conductor Cloud
Charlie opens by framing Conductor as a Mac app for running many coding agents in parallel across isolated copies of your codebase, then reviewing and merging the results. The big launch in this interview is Conductor Cloud: until now, closing your laptop killed the agent, but now workspaces can stay alive in the cloud and keep coding without you.
Fast growth, but a very human bottleneck
The company is still a small team, but they say revenue and usage have grown about 10x since January, with adoption ranging from indie hackers to engineers at large public companies. But Charlie is candid that the limit isn't just compute — it's your brain: right now he can only really hold three to five agents in working memory, which makes the next problem an interface problem, not just a model problem.
How two college teammates became co-founders
The origin story is unexpectedly personal: they met in college, where Charlie was a fifth-year, Jackson was a freshman, and Charlie taught him how to lift. Later they reconnected while Jackson was on Netflix's ML infra team and Charlie was doing engineering and growth at Replicate, and realized they loved building together enough to go all-in — with a shared desire to recreate the feeling of being on an elite ultimate frisbee team, just now against companies instead of UNCC.
The bad first idea, and the YC reset that mattered
Their original YC application was for AI that books reservations and even tennis courts for you — interesting, but by their own telling, a 'solution in search of a problem.' They spent the first month and a half of YC cycling through ideas every few days, until Aaron Harris gave two pieces of advice they still remember: focus on devtools, and put a poster of yourselves on the wall that says 'make something these guys want.'
Before Conductor, there was a failed 'thing after the IDE'
That advice pushed them toward AI coding, where they felt tools still weren't going far enough after Sonnet 3.5. They actually tried to build a post-IDE interface that let humans just tell AI what to do and review the output, but the models weren't ready yet — too much handholding, weak tool use, too many trips back into the IDE — so they retreated to something simpler and shipped a multi-model chat app called Chorus.
They discovered Conductor while building their own tools
The twist is that Conductor came out of the internal devtools they made to build Chorus faster. Charlie hacked together an MVP in about a week, Jackson reacted with some version of 'this is not what I asked for, but wow, this actually works,' and within roughly two to three weeks they had something in users' hands. The moment it clicked was hilariously tactile: assign one agent a task, hit Command-N, start another, then see the unread dot appear because the first one already finished.
What top engineers are actually doing with AI
From all their in-person user research and bike rides to meet engineers, one lesson stood out: the best users spend serious time on 'skills files' — markdown docs that encode React best practices, codebase quirks, and constraints agents should keep re-reading. At the same time, many of those engineers have surprisingly plain setups; Jackson compares over-optimized AI workflows to elaborate Vim setups that look impressive but don't always ship more work.
Their bigger vision: from agent tabs to AI org charts
Looking forward, they think models will get 10x to 100x smarter and behave more like coworkers, even if they still have an 'alien brain.' Charlie's metaphor is that future interfaces should make you feel like the CEO of a megacorp of AI employees — mostly operating at a high level, then drilling into details when needed — and he says both working memory and code review are still major bottlenecks, especially since software teams are still reviewing AI work with workflows that feel stuck in GitHub circa 2010.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
