{"slug": "experiential-labs-open-sources-a-router-that-learns-from-agent-traffic", "title": "Experiential Labs open-sources a router that learns from agent traffic", "summary": "Experiential Labs, a San Francisco startup in Y Combinator's Summer 2026 batch, open-sourced World Model Optimizer, a model router that learns from an AI agent's production traces to route requests to cheaper models, aiming to reduce reliance on rented frontier models. The tool, introduced in a July Hacker News post by co-founder Silen Naihin, uses K-nearest-neighbor to select models based on similar past work and can optimize for cost or quality, with confidence gates to avoid switching from the strongest known model. The company's gateway also records traces to evaluate alternatives and train specialized models, with customers owning the resulting weights.", "body_md": "# Experiential Labs open-sources a router that learns from agent traffic\n\n**The Y Combinator Summer 2026 startup is betting that production traces can push routine work onto cheaper models customers own.**\n\nBy [Ryan Merket](/author/ryan-merket)\n· Published\n\nPrimary source: [Hacker News](https://news.ycombinator.com/item?id=49063454)\n\n## Why it matters\n\nModel gateways are becoming control points for enterprise AI spending. Experiential Labs is betting that owning the traffic layer also creates the data needed to reduce dependence on rented frontier models.\n\n[Kion Fallah](https://kionfallah.com/?ref=runtimewire) and [Silen Naihin (@silennai)](https://x.com/silennai?ref=runtimewire) open-sourced a model router in July that uses an AI agent's production history to decide which model should handle each new request. Their larger bet is that the inference bill can eventually finance its own replacement.\n\nThe two founders are building [Experiential Labs](https://www.experientiallabs.ai/?ref=runtimewire), a San Francisco company in Y Combinator's Summer 2026 batch. Experiential Labs sits between an agent and the models it calls, offering one [OpenAI-compatible endpoint](/article/vlm-run-gateway-open-weight-ocr-models) for hosted providers, customer-owned API keys, local models and custom models. The gateway also records the traces needed to evaluate cheaper alternatives and train specialized models.\n\nFallah came to the problem from autonomous driving. He earned a machine learning Ph.D. at Georgia Tech, then led development of a mixed-reality simulator at Waabi for testing autonomous vehicles. His thesis at Experiential Labs carries that method into software agents: simulate the environment, replay difficult cases and test proposed changes before putting them into production.\n\nNaihin approached the same problem through open-source agents and continual learning. [Y Combinator's profile](https://www.ycombinator.com/companies/experiential-labs?ref=runtimewire) credits him with helping AutoGPT grow to about 160,000 GitHub stars. His [personal biography](https://www.silennai.com/?ref=runtimewire) says he later worked with the U.S. Department of Energy on AI for scientific discovery and spent two years running Stackwise, a YC Winter 2024 company.\n\nThose backgrounds explain why Experiential Labs is trying to build an improvement loop rather than another static proxy for model APIs.\n\n### A gateway with memory\n\nIn a [July Hacker News post](https://news.ycombinator.com/item?id=49063454&ref=runtimewire), Naihin introduced World Model Optimizer, an open-source tool that ingests agent traces and constructs text simulations of the tools and environments an agent encounters. It runs candidate models against those simulated tasks, embeds the requests and fits a K-nearest-neighbor router that selects a model based on similar past work.\n\nThe router can optimize for cost or quality and accounts for prompt caching when estimating the effective price of a request. Confidence gates keep it from switching away from the strongest known model when the available evidence is weak or when a request looks unlike the router's training examples.\n\nThat process gives Experiential Labs a path beyond ordinary provider routing. A conventional gateway can send a request to the cheapest available host, fail over during an outage or enforce a budget. Experiential Labs wants the gateway to learn which classes of work can move to cheaper open models and, eventually, to a specialized model trained on the customer's own traces.\n\nThe current [open-source repository](https://github.com/experientiallabs/experiential?ref=runtimewire) combines the gateway and optimization workflow. Developers can run the gateway locally, connect provider keys and build a router from OpenTelemetry traces. The repository had more than 380 GitHub stars at the time of the supplied research.\n\nExperiential Labs says customers will own the weights produced through its training workflow. That ownership pitch gives the founders a direct answer to the recurring complaint that companies spend heavily on frontier inference while accumulating little infrastructure they can operate independently.\n\n### The benchmarks still belong to Experiential Labs\n\nExperiential Labs reported [three self-run routing results in its July post](https://news.ycombinator.com/item?id=49063454&ref=runtimewire). On RouterBench, the router cut cost by 66.5% while performance fell 1.7%. On TauBench, Experiential Labs reported 44.5% lower cost and 6.3% higher performance. On Terminal Bench 2, Experiential Labs reported a 64% cost reduction and an 8% performance increase.\n\nThose figures measure configurations chosen and run by Experiential Labs. They are not independent validation, and they do not establish that a router trained on one customer's traces will deliver similar savings on another customer's workload.\n\nThe launch discussion exposed the more important proof gap. One commenter asked Experiential Labs to release a model trained with the method and show the improvement over its base model.\n\nExperiential Labs makes broader claims elsewhere, including up to 97% lower cost on some tasks and a guarantee of 50% savings at equal or better quality. Those claims remain company assertions. Experiential Labs has not published customer counts, revenue or the contract terms behind the guarantee.\n\n### The gateway is the wedge\n\nThe gateway market already has well-funded incumbents. OpenRouter provides unified access, routing and failover across model providers. In [its May 2026 financing announcement](https://openrouter.ai/blog/announcements/series-b/?ref=runtimewire), OpenRouter said it raised a $113 million Series B and was processing 25 trillion tokens a week for more than 8 million developers.\n\n[LiteLLM](https://docs.litellm.ai/?ref=runtimewire) offers a self-hosted open-source gateway, while [Vercel AI Gateway](https://vercel.com/ai-gateway?ref=runtimewire) bundles hundreds of models behind one endpoint with budgets, observability and provider failover. [Portkey](https://new.portkey.ai/announcements/portkey-has-raised-15m-in-series-a-funding?ref=runtimewire) raised a $15 million Series A in February to expand its enterprise gateway and governance products.\n\nExperiential Labs cannot win by offering another menu of models. Its stated distinction is an optimization layer that uses agent traces and simulations to build a customer-specific router or specialized model. If that loop works, each production request becomes evidence for moving future work away from expensive frontier capacity.\n\nOpen source helps lower the trust barrier. Developers can inspect and run the routing code locally before sending production traces to a hosted service. Experiential Labs can then sell managed training, serving and continual optimization to customers that do not want to operate the pipeline themselves.\n\nNaihin says Experiential Labs has backing from Y Combinator and angels affiliated with Meta, Google DeepMind, Nvidia and Waabi. Experiential Labs has not announced the amount raised or named the individual angel investors.\n\nThe founders have picked a problem with clear economics: agents generate repetitive work, frontier models charge for every repetition, and production traces contain evidence about which requests need expensive reasoning. Experiential Labs now has to prove that its simulations reflect real environments closely enough to make routing and training decisions customers can trust. The open-source release gives developers the machinery to test that claim rather than taking another AI cost-saving promise at face value.", "url": "https://wpnews.pro/news/experiential-labs-open-sources-a-router-that-learns-from-agent-traffic", "canonical_source": "https://runtimewire.com/article/experiential-labs-open-source-ai-router-agent-traffic", "published_at": "2026-08-27 22:17:21+00:00", "updated_at": "2026-08-27 22:48:20.946592+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-infrastructure", "ai-tools", "ai-startups"], "entities": ["Experiential Labs", "Y Combinator", "World Model Optimizer", "Silen Naihin", "Kion Fallah", "AutoGPT", "Stackwise", "Waabi"], "alternates": {"html": "https://wpnews.pro/news/experiential-labs-open-sources-a-router-that-learns-from-agent-traffic", "markdown": "https://wpnews.pro/news/experiential-labs-open-sources-a-router-that-learns-from-agent-traffic.md", "text": "https://wpnews.pro/news/experiential-labs-open-sources-a-router-that-learns-from-agent-traffic.txt", "jsonld": "https://wpnews.pro/news/experiential-labs-open-sources-a-router-that-learns-from-agent-traffic.jsonld"}}