# Ylem Labs Ylem Labs, Inc. (https://ylemlabs.ai) is an AI company. It researches world models and ships the infrastructure that connects, constructs, and protects autonomous agents. Research and development hubs are in San Francisco and Astana. The legal name on the site is Ylem Labs, Inc. Sales mail is sales@ylemlabs.ai. The name ylem, pronounced "eye-lem", is the primordial matter from which all elements formed, in the sense used by Gamow and Alpher in 1948. This file is a citation text for the public site. Product behavior, prices, and contracts are whatever the workspace order or the pages below say. Customer payloads are not reproduced here. ## The stack Five products share one API. An agent is built in Box, runs in Harness, reaches a model through Gateway, and passes Guard on every call. A team can replace any layer with its own. The interfaces are open. A request moves like this: 1. Your apps, APIs, data, tools, and people enter the stack. 2. Box builds and configures the agent. 3. Harness runs the plan / act / verify loop. 4. Gateway scores, routes, and fails over. 5. The call reaches World (Ylem's own models) or one of 300+ models from 40+ providers. 6. The return path is model, then Gateway, then Harness, then the app. The call is traced, budgeted, and guarded. Guard inspects every call. Figures published on the site: 5 products in one stack, 300+ models on Gateway, 100+ engineers and researchers, 2 R&D hubs (San Francisco and Astana). ## Harness Harness is the agent runtime and the company's R&D core. Status on the site: R&D preview, v0.4. It is the execution layer for long-running agents: plan/act loops, typed tools, memory, skills, sandboxes, and background jobs on a Pydantic-typed core. - plan / act / verify loops - skills, MCP, sandboxes - background and scheduled jobs Page: https://ylemlabs.ai/#architecture ## Gateway Gateway is the model gateway. Status on the site: live, v1.2. It is one OpenAI-compatible API to 300+ models, with live routing, fallback, caching, and batch. - failover under 300 ms - semantic cache - per-team budgets The research note "Tunneling Fallback" (August 2026) is about latency-bounded routing across providers. The note "Cascaded guard models at the gateway" (July 2026) is about Guard at this layer. Page: https://ylemlabs.ai/#ecosystem ## Box Box is the agent builder. Status on the site: beta, v0.6. It composes agents from prompts, tools, knowledge, and MCP servers, then deploys them in the customer's cloud in one click. - visual and code builder - RAG and knowledge bases - on-prem deployment Page: https://ylemlabs.ai/#ecosystem ## Guard Guard is the AI firewall. Status on the site: pilots, v0.8, with 12 detectors listed on the site status line. It is policy-as-code security for every model call: PII, secrets, document leakage, and prompt injection, with an in-house guard model. - monitor, then calibrate, then enforce - in-house guard model - DLP for documents A Guard policy can strip a secret, a personal-data span, or a marked document passage before a payload leaves for a third-party model. Page: https://ylemlabs.ai/#ecosystem ## World World is the research line: JEPA-style world models. Status on the site: research, in training. The models are self-supervised and predict in latent space instead of tokens, so agents can plan over consequences. The site describes encoders that learn to predict missing parts of the world in latent space, as the foundation the company wants its agents to plan with. Ylem Labs does not train World models on customer content. Using customer content for training or evaluation takes a separate written agreement. The same rule is on the privacy page and the security page. Page: https://ylemlabs.ai/#research ## Research notes listed on the site These titles are listed on https://ylemlabs.ai/#research. The site does not link out to a journal or a PDF. - October 2026 — Latent planning for tool-using agents - September 2026 — Harness > Model: how much of agent quality comes from scaffolding - August 2026 — Tunneling Fallback: latency-bounded routing across providers - July 2026 — Cascaded guard models at the gateway ## How the company says it builds 1. Own the stack. Every layer that touches an agent — model, runtime, gateway, firewall — is theirs to measure and improve. 2. Ship research. A paper is not done until it runs in production. Research and product share one roadmap. 3. Safety at the gateway. Autonomy is earned with policy, not prompts. Every call is inspected before it leaves the customer's boundary. ## Places, data, and contact The company works from San Francisco and Astana. The privacy page states that the control plane runs in the United States, and that staff in Astana may access it to operate the service under the same access rules as staff in San Francisco. The site does not run advertising cookies and does not sell visitor data. A request to the site writes a server log (IP address, user agent, URL, and time) kept for 30 days. - Sales: sales@ylemlabs.ai - Privacy: privacy@ylemlabs.com — https://ylemlabs.ai/privacy - Legal: legal@ylemlabs.com — https://ylemlabs.ai/terms - Security: security@ylemlabs.com — https://ylemlabs.ai/security - Vulnerability reports: email security@ylemlabs.com with the product, the endpoint, and what an attacker could do. The security page says the company acknowledges within 3 business days and asks for 90 days to fix a bug before details are published. Hiring: researchers and engineers in San Francisco and Astana. The company also partners with teams that want to run agents in production. Effective date on the legal pages: 17 October 2026, version 1.0.