AI builders are moving from factory automation to orchestrated, human-led agent teams

The pitch. Charlie Holtz, Conductor's co-founder, opens with principles from high-performing AI builders, not a product walkthrough. The demo drives the point: his desktop app manages coding agents in one interface, and the release moves tasks from local Git worktrees into cloud sandboxes so agents keep running when a laptop closes. Teams watch colleagues and agents work in shared spaces, and agents can spawn workspaces through an API. The product is the argument.

The workflow. Holtz says fastest builders aren't pure token maxers. First principle: stay near the frontier. Try new tools on day one; waiting for your social graph means being three to six months behind. Conductor emerged from Claude Code usage, after the team cloned its repo five times, discovered worktrees, and built an internal tool. Then: don't beat the market. If a workflow is becoming the default, wait for the model to absorb it. Invest only where you have alpha.

The safeguards. The phrase likely to travel is “slop free zones”: parts of the codebase require strict human review, while other parts stay loose. Conductor learned this after rewriting its app twice. In CI, any change to the migrations file requires a human. Anything in Slack is human-written by definition; docs and CLAUDE.md/skills files get disproportionate care because agents load them every session. The governance pattern: pick small high-risk paths, keep them human, spend effort on context.

The context machine. Feed the beast is the centralization play. Every Slack message, Discord bug report, and meeting recording is captured into a Postgres-backed internal agent, the CIA, so agents query company-specific context with SQL instead of guessing. This is the most concrete idea and the one with staying power: agent usefulness depends on organizational memory being machine-readable. It also creates obvious privacy and retention questions that Holtz does not address.

The trade-off. The orchestras, not factories closing is more philosophy than evidence. Holtz rejects the factory framing because feature factories failed and managers should not be pushing buttons. He wants human flow, with agents as an orchestra the leader can zoom in and out of. That is a sharp counter to swarm rhetoric, but an aesthetic position, not a measured result. There are no benchmarks, team-size numbers, or before/after comparisons; the collaboration claim is unsupported.

The signal. The actionable core is not the metaphor or product. It is three choices: keep critical code paths under human control, route organizational context into agent-accessible storage, and give agents an environment that outlives a laptop. Those fit what teams need before scaling agent use. The missing piece is evidence: one founder's observation, one product demo, and a lot of normative language. Steal the patterns, wait on the platform.