The agents already orchestrate. That is what they do: a coding agent takes an instruction, breaks it into steps, calls tools, and drives the work.
So the interesting question is not who orchestrates the code. It is who orchestrates the orchestrators.
Agent FM is going for that seat, and the seat has a property worth understanding before anything else about the company.
What they built
Agent FM is a desktop app that gives an engineer one group chat to hear and steer their coding agents.
Instead of switching between separate Claude Code and Codex sessions and reading walls of output to work out what each one is doing, the agents report into a shared chat. They post progress, blockers, assumptions they have made and decisions they have taken. The engineer gets those updates as live text and audio, and replies straight from the chat.
The founders are Aniket Deshpande, chief executive, previously in applied AI at Meta Superintelligence Labs, and Mugdhaa Patankar, chief technology officer, previously at Snowflake, Twitch and AWS. It is based in San Francisco.
Who else is in this field
Everyone is building coding agents and almost nobody is building the place you stand above them. Cursor, GitHub Copilot, Windsurf, Cognition and Augment Code all make the agent itself better. OpenHands and Aider took the open source route. Anthropic and OpenAI ship their own coding agents and their own surfaces around them. Linear and Jira are where human engineering work is tracked, Slack is where humans coordinate, and PagerDuty solved the related problem of getting a person's attention when a machine needs it.
Every one of those is a thing to be orchestrated, or a place humans talk to each other. None of them is a neutral floor above the agents.
The point: the model companies structurally cannot take this seat
This is the argument, and it is the whole reason the position exists.
An aggregator is only useful if it is complete. A cockpit that shows you three of your four agents is not a cockpit, it is a fourth window. So whoever sits on top has to be willing to give every agent equal billing, including the ones that compete with each other.
A model company cannot do that. Anthropic will never build a desktop coworker that gives Codex equal standing next to Claude Code, and OpenAI will never build one that gives Claude equal standing. Their own surfaces exist to keep you inside their own model.
That is not a failure of imagination or a gap they will close next quarter. It is what they are for. Every incentive they have points at their agent being the one you use, and an honest aggregator's incentive points the other way, at being complete regardless of who made what.
Which means the seat above the agents is available only to somebody who sells no model. That is a very short list, and it is the strongest thing Agent FM has going for it. Not the chat, not the audio, but the fact that the largest companies in the field are disqualified from the position by their own business model.
The features are the mix, not the point
It is worth being straight about the product itself.
A single shared chat window is a reasonable idea. Audio updates are a genuinely clever touch, since a supervising engineer's eyes are occupied by their own work and hearing is the free channel. Deciding what an agent should bother telling you, rather than dumping a log, is real craft.
But none of that is defensible on its own. Any of it can be copied in a fortnight by anyone who decides it matters. These are the mix, and a mix either catches or it does not.
What cannot be copied is the seat. Features decide whether you get to sit in it. The seat is what makes sitting in it worth anything.
So the right way to read the product is as the audition rather than the asset. It has to be good enough to become the thing engineers download when they get their third agent and lose track of what all of them are doing.
What you actually win at the top of the pile
Here is the part worth carrying out of this file, and it generalises well beyond this company.
Being the thing everybody downloads is worth more than the downloading. When a tool becomes the one people reach for during a moment, the company acquires something that outlasts the moment: the attention, the relationships, and the direct line to the people who happen to be building the most consequential software of this period.
That audience does not expire when the product cycle turns. The engineers who install an agent cockpit in 2026 are the people who will be running AI infrastructure for the next twenty years. Reaching them once, on good terms, while they actually want something you have, is a genuinely scarce thing.
There are thousands of startups and very few of them ever get a window where the right people are actively looking for what they built. A window like that is worth taking even if it is short, because the relationships it creates outlive whatever created them.
That reframes how to judge a company in this position. The question is not only whether the group chat is the permanent answer to multi-agent supervision. It is whether Agent FM can be the obvious thing to install during the period when every engineer is working out how to run four agents at once, because whoever is obvious then is holding something durable afterwards.
Plenty of well-known companies have run exactly this sequence: win a moment, earn an audience, then build the next thing for the audience you already have. It is an ordinary and honourable path, and it is available to almost nobody, because the moment has to exist and you have to be standing in it.
Why it is bigger than it sounds
The short description is a group chat for coding agents. The position is a seat the biggest companies in AI are not allowed to take.
- An aggregator has to be complete, and the labs cannot be. A model company will never give a competitor's agent equal billing in its own product, which leaves the neutral floor above the agents to someone who sells no model.
- The features are the audition, not the asset. The chat and the audio can be copied in a fortnight. The seat cannot, and the features exist to win it.
- The audience outlives the product. Being the thing engineers install during this particular moment is a line to the people building AI infrastructure for decades, and very few companies ever get such a window.
What to watch
The number that matters is how many different agent makers are represented in a typical customer's setup. If almost everyone is running one vendor's agents, this is a companion app for that vendor and the vendor can absorb it. If real users routinely have agents from two or three different companies in the same chat, the neutrality argument is live, the seat is real, and nobody with a model to sell can follow.
The second is whether people leave it running. A supervision tool that engineers mute after a week has failed at the craft part, and the mute rate will tell the founders more than any other number they could collect.
The third is growth that costs nothing, because that is what a moment looks like from inside. If engineers are arriving because another engineer told them to, Agent FM is standing in the window. What it does with that is the next question, and a much better problem to have.