YC Just Open-Sourced Its Multiplayer Harness

BBetter Stack
Computing/SoftwareSmall Business/StartupsInternet Technology

Transcript

00:00:00do you remember when a new javascript framework would be released every week well the same thing
00:00:04is happening again but with agent harnesses and this time the guys that usually throw money around
00:00:09have actually picked up a keyboard and wrote some code well actually claude probably did it but
00:00:14startup accelerator y combinator have released their own agent harness qm and they call it a
00:00:19multiplayer agent harness for work in slack and on the web and it does have some interesting
00:00:24features plus it's specifically designed for startups so if you run a company that makes zero
00:00:29profit then this may be the tool for you so what actually is qm if you've read through the github
00:00:39repo you've probably come away with more questions and answers because it does a pretty terrible job
00:00:44of explaining itself the word harness makes it sound like another clawed code something you install on
00:00:49your laptop and run in your terminal it's actually a full stack web application that companies
00:00:54deploy onto their own infrastructure and everyone in the org gets their own isolated workspace on it
00:01:00underneath that you've got a headless backend running the agent loops and postgres holding
00:01:05the session history so none of it lives on your machine it plugs into slack so people can use it
00:01:10wherever they already work or you can drive it directly from the web ui so if you want to set
00:01:15it up once and forget about it you can do that and if you want to stay up to date with the latest
00:01:19tools then subscribe to better stack so why would you even use this well let's break it down qm runs
00:01:26in the cloud not on your machine that means the agent and the tool calls all happen off device and you
00:01:32can access it on any device that connects to the web and every workspace gets its own container with its
00:01:38own storage attached so anything the agent installs stays installed get it to set up a python environment
00:01:44on monday and it's still there on friday along with the files and anything it cloned so each workspace
00:01:50is a real machine that keeps its own state next up is shared context you can create a project and add
00:01:56team members so you can share memory files apps and connectors between the organization that means you
00:02:02can set things up once and everyone benefits by default everyone gets a personal space which is effectively
00:02:09a project set to private so you can choose to set up only things you have access to and safely configure
00:02:15things like credentials without sharing them however if you do want multiple team members to have access
00:02:20to the same tools files or credentials that is also possible the web ui actually does a terrible job of
00:02:26telling you which context you're in with basically nothing on screen letting you know where you're configuring
00:02:32things against you can see here i click my shared project and in files i get a small pill to tell
00:02:37me i'm uploading to that project but in other tabs i see no indicator at all i would expect to see a
00:02:43persistent ui element to tell me which context i'm in but i'm sure that can be added soon the next benefit
00:02:49is harnesses if you do really like the way a certain harness works like a clawed code for example
00:02:54you can configure that to run inside qm you can see here the default harness is pi but i can switch to clawed
00:03:01and i'll be running the exact same clawed code harness i'm used to but in a cloud environment
00:03:06getting clawed code into that list actually took a little bit of digging and admin decides which harnesses
00:03:11everyone is allowed to pick and out of the box you only get the one your deployment booted with
00:03:16so my drop down started with just pi and nothing else there's no control for it anywhere in the admin
00:03:21panel so i had to set it with an api call that's a put at the harnesses endpoint with the ids of the
00:03:27two harnesses i want and after refresh they both show up the next benefit is that it has a really tiny
00:03:32tool surface now this is something that i really liked the agent only gets seven tools execute read
00:03:38write publish memory history and background everything else it does is just a shell command
00:03:43inside its own sandbox so instead of wiring up a pile of mcp servers hoping the model picks the right
00:03:49one you can just hand it a machine and let it work the way you would it's the same approach that
00:03:53claude code takes and it's nice to see that this is built at the core of the design so how do you let
00:03:58an agent loose on company data without it going badly wrong there's an org wide setting for that
00:04:03with three levels on strict just about every tall call pauses and waits for a human to approve it auto
00:04:09is the default and it runs a classifier over external data and tool results before they reach the model
00:04:16and on dangerous there's no screening and nothing pauses a team can tighten that for their own scope but
00:04:21they can't loosen it past what the org picked there's also a command policy that applies whatever posture
00:04:27you're on so things like recursive deletes and destructive sql get blocked even on dangerous
00:04:33mode you'd want something like that before putting an agent in front of the whole company
00:04:36and it's not something you get when everyone runs their own agents locally overall i can see qm
00:04:42being a powerful tool for teams having shared access to company documentation because as an admin you can
00:04:47configure fine-grained access to external tools through credentials choose to include files or
00:04:52memory on how to run various processes you can even create shared skills that anyone in your project can
00:04:58access so rather than isolated setups like we have today with our cli tools this allows organizations
00:05:04to benefit from everyone's hard work personally i think it's worth a look it's already gained 13 000
00:05:09stars on github but i would say there's still much to be done about the ui and ux to be fair to the
00:05:14developers the project is only two weeks old so it literally is not mature however the potential here
00:05:18is huge it's open source and you can control and tweak absolutely everything including changing the
00:05:24source code if needed so you can check it out on the github repo linked in the description
00:05:29and if you enjoyed that don't forget to subscribe to better stack and i'll see you in the next one

Key Takeaway

Y Combinator's open-source qm harness provides companies with a cloud-based, multi-tenant agent platform featuring persistent container storage, shared project contexts, and configurable safety guardrails.

Highlights

  • Y Combinator released an open-source multiplayer agent harness called qm that runs on company infrastructure rather than local developer laptops.

  • The qm repository gained 13,000 stars on GitHub within two weeks of its launch.

  • Each workspace operates inside an isolated container with persistent storage attached so installed environments and files remain intact across sessions.

  • The agent utilizes a minimalist tool surface consisting of only seven core tools: execute, read, write, publish, memory, history, and background.

  • Three security posture levels control agent permissions: strict pauses for human approval on tool calls, auto runs a classifier over external data, and dangerous disables screening.

Timeline

Qm Architecture and Infrastructure

  • Qm is a full-stack web application that companies deploy on their own infrastructure.
  • A headless backend executes agent loops while postgres stores session history.
  • The system integrates with slack and provides a web ui for direct control.

Startup accelerator y combinator released an agent harness named qm designed for multiplayer work in slack and on the web. Unlike terminal-based local tools, it operates as a full-stack web application deployed directly on company servers, ensuring no data lives on individual user machines.

Cloud Workspaces and Shared Context

  • Every workspace runs in the cloud with attached storage that preserves installed environments and files.
  • Shared projects allow teams to distribute memory, files, apps, and connectors across the organization.
  • Personal spaces default to private configurations for secure credential management.

Workspaces function as real machines that maintain their state over time, meaning python environments and cloned files persist from monday to friday. Team members can collaborate within shared projects or utilize private personal spaces to configure credentials without exposing them broadly.

Harness Selection and Tool Surface

  • Administrators can configure alternative harnesses like claude code to run inside qm.
  • The agent relies on a minimal set of seven native tools.
  • All other complex operations execute via standard shell commands inside the sandbox.

Users can swap the default pi harness for other options like claude code through api configurations. The agent operates with a restricted tool surface of only seven functions, relying on sandbox shell commands instead of complex multi-server mcp setups.

Security Guardrails and Deployment Potential

  • Organization-wide safety settings offer strict, auto, and dangerous execution modes.
  • Command policies block destructive actions such as recursive deletes and sql drops even in dangerous mode.
  • The project provides a foundation for team-wide agent collaboration despite current ui limitations.

Administrative settings govern tool call approvals and external data filtering to prevent unintended agent behavior. Command-level restrictions enforce safety barriers against catastrophic actions, positioning the open-source platform as a scalable tool for organizational documentation and workflow automation.

Community Posts

View all posts