Conductor: how to conduct an orchestra of agents (Series #AI4SDLC)
In my review of The Ignorant Maestro, I wrote that I have long been interested in how conductors lead ensembles and what software development can learn from them. Now agents are being invited to join the people in the orchestra.
That is the metaphor Charlie Holtz, co-founder and CEO of Conductor, uses in his talk “Orchestras, Not Factories.” He spoke at AI Engineer World’s Fair on June 30, 2026; the standalone recording came out on September 27. Conductor is an app for managing multiple coding agents and working on tasks with colleagues. Holtz wants people to set the direction, get into the details, and enjoy building a product, while agents take on more and more of the execution. Behind the appealing metaphor are several practical rules.
🔸 Try new things without getting stuck in configuration In his view, you should refine your workflow where the product or team has particular needs. Major tool developers will eventually build in general improvements themselves. It is a good reason to check how much time the next tweak takes and how much it actually saves. 🔸 Define areas that require human review Conductor calls these slop-free zones. Changes to migrations, for example, need a human reviewer. The team also prepares documentation, agent instructions, and skills especially carefully. I like the selectivity here: identify in advance where mistakes cost more. A poor instruction, for instance, can affect many subsequent tasks. 🔸 Make the team’s knowledge accessible to agents Their Conductor Internal Agent collects Slack messages, Discord requests, and meeting recordings in PostgreSQL. Agents can search that information through SQL. It is a useful idea, but I would think separately about distinguishing a discussion from an accepted decision, identifying outdated information, and respecting access rights. A human message can be wrong too.
Holtz then demonstrates collaboration in the cloud. There has already been a development beyond the talk: Conductor Cloud launched on July 30. Agents work in separate environments and continue tasks after the laptop is closed; colleagues can join a shared workspace. An iOS app followed on October 2.
Comparing this with other approaches, I would describe their emphases this way. Tereza Tížková from Factory shows the organization of execution and verification in more detail. Dex Horthy from HumanLayer focuses on the engineer’s involvement in design and retaining an understanding of the code. Holtz focuses on how a person and a team manage multiple executors. These approaches can work together.
Conductor sells exactly this kind of working environment, so its commercial interest is clear. The talk contains no comparative speed measurements: it presents a founder’s observations and a demonstration of his product.
This brings us back to Zack Proser’s discussion of human attention. Conducting an orchestra can be exhausting too. If you spend all day switching between agents, repeating explanations, and sorting out unfinished work, a new metaphor changes little. I am more interested in which recurring interventions this process lets us eliminate.
Three practical steps follow: 1️⃣ For engineers Choose several areas where human review is mandatory, and reread the instructions the agent receives every time it starts. 2️⃣ For platform teams Check whether an agent can find the current decision, start the environment, and continue a task without constant manual hints. 3️⃣ For managers Compare the pilot with the usual process: how much human time goes into task definition, switching, review, and rework? At comparable quality, this will help estimate how much attention working with multiple agents frees up.
#AI4SDLC #AI #Agents #Engineering #Management