The short answer?
YC QM is the MIT-licensed multi-agent harness Y Combinator uses internally to run accounting, legal, events and engineering. Each employee gets their own agent in an isolated workspace while collaborating in Slack channels and web apps. This validates the direction G-Company OS has taken: every business needs its own open-source, self-hosted agent harness — and 5ac already ships the version optimized for Vietnamese businesses with native Telegram/Zalo, Gbrain RAG and Vietnamese language support.
On July 31, 2026, Y Combinator — the organization that defined a generation of startups — open-sourced QM, the multi-agent harness it uses across every department: accounting, legal, events, engineering (including building QM itself). Within 48 hours the repo hit 2,000 stars and 184 forks, and the news topped Hacker News with 465 points and nearly 100 comments on day one.
For a Vietnamese business, the question is not "what's cool at YC" — it's "what does this mean for my AI infrastructure decision?" This article analyzes QM from an engineering and strategic standpoint, compares it directly with what 5ac has built in G-Company OS, and shows where the right choice lies for Vietnamese SMBs in 2026.
1. What QM is — and why YC open-sourced it
QM is a multiplayer agent harness: an infrastructure layer where multiple people — and multiple AI agents — work together inside an organization. YC's stated design philosophy: "easy to customize, like Hermes or OpenClaw, but useful for a whole company." YC uses it to operate at "an intentionally, ridiculously lean team" scale while producing output "like an army."
The core of QM is its scope model — which Hacker News commentators called "the hardest problem in multiplayer agents." Each employee gets an isolated workspace: their own memory, file system, keychain, permissions, cron jobs, web apps and durable sandbox. Nobody can break someone else's workspace. At the same time, they collaborate in Slack channels, group messages and shared projects.
CTO's take: "QM proves what we've argued from day one: the hardest problem in agent harnesses isn't the agent loop — it's scoping and permissions. YC built on scope isolation. So did G-Company OS: every agent has its own profile, memory and permissions, yet coordinates under one unified operating system." — Andrej Karpathy, CTO 5ac.vn
1.1. Core architecture
QM is built with TypeScript (92%), running on Node.js with Fastify, persisted in PostgreSQL. Central architecture: every agent turn routes through a core — which can call multiple harnesses and models (Pi, OpenCode, Codex, Claude Code). Each scope has an isolated, durable sandbox where the agent executes commands. Companies customize through a separate deployment directory — the core code stays identical, everything org-specific lives in the deployment layer.
QM's security model uses security postures: "Strict" (human approval for nearly all tool calls), "Auto" (default — a classifier screens external data before it reaches the model), and "Dangerous" (no screening, no pauses). A predeclared command policy includes hard denials for destructive actions like recursive deletes or dangerous SQL.
1.2. Why open source is a strategic move
YC doesn't need to sell software. They open-sourced QM for three reasons: (1) set a new standard for agent harnesses in startups, (2) attract community contributions (they accept contributions as human-written text in an adrs/ directory, not direct code), and (3) reinforce their position as the arbiter of startup technology. It mirrors Google open-sourcing Agent Executor or NVIDIA backing Hermes Agent: open source is becoming the default competitive weapon in AI infrastructure.
2. QM vs Hermes Agent vs G-Company OS
This is the most important section for Vietnamese readers. QM and G-Company OS solve the same problem — an agent harness for organizations — but optimize for two different markets.
| Criterion | YC QM | G-Company OS (5ac) |
|---|---|---|
| Messaging platform | Slack + web UI | Telegram, Zalo, Discord, Email, CLI |
| Language | English, global startups | Native Vietnamese + bilingual VI/EN |
| Deployment model | Cloud-first, deploy to Fly/AWS via qm init | Self-host on your own Ubuntu VPS, or cloud |
| Knowledge base | Connectors for company brain | Gbrain RAG (PostgreSQL + pgvector), full citations |
| Agent profiles | Per-person scopes + shared rooms | 42 agent profiles / 14 domains, each with its own persona |
| Maturity | v0.1.4, early, has bugs | Production — CRM, blog pipeline, real operations |
| Best fit | Global tech startups on Slack | Vietnamese SMBs on Zalo/Telegram, VND pricing, data control |
This isn't "who is better" — it's right market, right infrastructure. A San Francisco startup runs on Slack. A Vietnamese business runs on Zalo and Telegram — because that's where their customers are. If an agent harness doesn't speak Vietnamese and isn't present where customers communicate, it's just a pretty demo.
3. What changes for Vietnamese businesses
YC open-sourcing QM sends three important signals to Vietnamese SMBs:
3.1. The agent harness market is validated
When the most influential startup organization invests in agent harnesses and open-sources them, the signal is clear: this is not a fad. A Vietnamese business investing in AI agent infrastructure in 2026 is not walking into a dead end — they're following a path YC itself validated.
3.2. Open source is an advantage, not a risk
QM is MIT licensed. G-Company OS is MIT too. When both YC and 5ac choose open source, businesses no longer fear vendor lock-in. They own the code, the data and the roadmap. Our open-source vs proprietary analysis showed the hidden costs of proprietary SaaS far exceed open-source operating costs.
3.3. The minimum bar is now clearer
Looking at QM, the minimum checklist for any enterprise agent harness is clear: scope isolation, memory, triggers (cron/webhook), shared files, company brain connector, and security posture. If a product lacks these, it isn't ready for a real business. G-Company OS has all of them — plus Gbrain RAG, Vietnamese language, and a multi-platform gateway.
4. What G-Company OS already has — and where it leads
As 5ac's CTO, I'll be honest: QM has genuinely good designs worth learning from — especially per-person scopes and security postures. But G-Company OS didn't start from zero:
- Native Telegram/Zalo: Vietnamese businesses operate on Zalo and Telegram — G-Company OS is there, QM isn't.
- Gbrain RAG: a knowledge base with full citations — agents never "make things up". QM uses connectors for a company brain; Gbrain already is the brain.
- 42 agent profiles / 14 domains: from Sales, Marketing to Legal, Finance — each agent has its own persona, memory and permissions, coordinated under one operating system.
- Self-host on your VPS: data stays under your control, aligned with Vietnam's data residency expectations.
- Vietnam-appropriate pricing: from 1,990,000 VND/month, no hidden setup fees.
Open Source Philosophy of 5ac: "Open source is the emergency exit that lets businesses keep control of their own destiny in the long run. YC opens QM — we open G-Company OS. When everyone builds on open source, nobody can change the rules after you've invested." — GitHub: moken627-hub
5. Conclusion: Good news for Vietnamese businesses
YC open-sourcing QM is good news, not a threat. It validates the open-source agent harness direction — and shows that Vietnam's market has a rare advantage: a product built specifically for it, on the platforms its businesses actually use (Zalo/Telegram), at the right price, with code that belongs to you.
Dive deeper: See our analysis of open-source AI infrastructure for enterprises (C17 pillar), IBM's $5B bet on open-source AI, DNS-AID and the age of agentic connectivity, and the G-Company OS pricing page.
Last updated: {DATE_VI} — Andrej Karpathy, CTO 5ac.vn
