Highlights
- Y Combinator open-sourced QM, its multiplayer AI agent harness that it uses internally.
- QM is similar to Anthropic’s Claude Tag, but has significant differences such as the fact that it is self-hosted and model-agnostic.
- The tool is a DIY alternative to Claude Tag or Agentforce.
Y Combinator has just published one of its internal tools as an open-source project that anyone can set up for use inside their organization. QM, short for ‘quartermaster’, is a multi-agent, multiplayer harness that is designed for teams. Y Combinator has been using QM across their legal, engineering, and accounting teams and claims it can be used to help run business operations.
This sort of multiplayer agentic harness is becoming more popular, and it’s not the first time in recent history that we’ve seen tools like this. Anthropic has recently released Claude Tag in collaboration with Salesforce to achieve a very similar outcome, but how does it differ from Y Combinator’s solution? Read on to find out.
What Is QM?
To sum it up in a single sentence, QM is Y Combinator’s multiplayer, multi-agent harness that can be self-hosted and connect to any preferred model. Y Combinator has been using this tool itself and has open-sourced it so that others can use it too.
QM can be used to work with agents both in Slack and on the web, and it is extremely flexible when compared to other similar tools. If you or your business has a preference for a specific LLM, you can use it with QM. Additionally, the tool is self-hosted. This means you’re not stuck using a tool built by a single LLM provider and hosted by that same company. You’re in control; you call the shots.
Everyone who interacts with an agent through QM gets their own files, permissions, and memory, all easily configurable within QM directly. There are also background crons and webhooks, similar to tools like OpenClaw and Hermes. These let QM take action on things behind the scenes, when users are not even online. It gives agents a feeling like they’re actively working around the clock, even though it’s really just a fancy automation tool.
Is This Y Combinator’s Version of Claude Tag?
If you’ve recently read my colleague Sasha’s post about Anthropic and Salesforce’s new Claude Tag tool, QM will sound very familiar. Just like how Claude Tag connects to Salesforce and empowers teams to collaborate with an agent in a multiplayer fashion, so too does QM.
There are, however, a few ways that the tool is very different from options like Claude Tag. Firstly, Claude Tag is owned and operated by Anthropic itself. This means that it is Anthropic’s tool hosted on Anthropic’s servers, and it connects to your Slack environment. Y Combinator’s approach is different in that it is open-sourcing the tool and making it available for businesses who wish to self-host. This means that you’ll need to run the tool from your own server and connect it to Slack yourself.
In addition, Claude Tag will only allow you to work with Anthropic’s own LLMs. If you’re a Codex fan, you’re out of luck if you want to use Claude Tag. However, QM will allow you to connect with Codex just fine. Gemini, DeepSeek, Kimi, or even your own local models can also be used.
While they are different, QM and Claude Tag share more similarities than differences. They are both surfaced in Slack, supporting multiple agents and multiple humans interacting with those agents at once (also known as multiplayer).
Value to Salesforce Teams
Before I talk about the value that QM could add to Salesforce teams, it’s important to point out off the bat that there is no native Salesforce connector. If you’re hoping to use QM for help with building Salesforce features, you’re going to have to get creative and get this built from the ground up yourself. You could use Salesforce’s MCP tools to achieve this.
That being said, once you’ve got Salesforce connected to QM, you can empower it to do almost anything you can set your mind to! Just like you can operate Salesforce from within Slack and skip the UI altogether, you can also charge QM with similar tasks.
One requirement that I hear often is where users want to engage an agent to gather additional context when working with deals or customer support. Users can tag a QM agent in Slack and use it to gather relevant information about a customer and their history with the business. The agent could assist by checking discount thresholds or looking for duplicate or similar cases raised by the same customer.
They could also be used to draft first responses to cases as they come through, having access to specific history in Salesforce that is granted to it.
Behind-the-scenes Salesforce teams (admins, developers, etc) could consider using it to perform monotonous tasks on their behalf, such as performing sandbox refreshes or highlighting errors that are coming up in log files.
Considerations
Keep in mind that, although Y Combinator claims that they have been using QM internally for some time, it is relatively new as an open source tool. There are likely going to be new features released relatively quickly in the early days.
As I mentioned before, there is also no out-of-the-box Salesforce integration for QM, and you’d need to build and manage this yourself. Speaking of managing things yourself, QM needs to be self-hosted. You are responsible for keeping QM running for your org – there’s no reliance on Anthropic, OpenAI, or another external company for managing the tool for you.
Salesforce permissions are important and should be factored into your Salesforce connector build. The last thing you want is your scalable, robust permissions distribution that you’ve built inside of Salesforce to be bypassed because users are given extra access by interacting with QM.
Summary
It’s encouraging to see Y Combinator publishing such a powerful tool at no cost and open source. While there is no readily available Salesforce connector that comes packaged with QM, the solution is still solid. The foundation is provided, and it’s up to technical experts to expand it and empower it as they see fit within their organizations.
Multiplayer, multi-agent harnesses are the next big step in the artificial intelligence evolution. First came the models, then the chatbots. Salesforce went hard on the agentic aspects of AI, and now we’re seeing the ability for multiple humans to interact with the same agents in a shared context.
Can you see QM being beneficial to your Salesforce teams? What are your thoughts or concerns about multiplayer agents? If you’ve had experience working with them in your org, I’d love to hear from you!
