Osmosis Wants Your AI Agent to Actually Work

Most AI demos look like magic. The trouble starts the week after, when the agent that dazzled in the pitch meeting quietly fumbles a multi-step task in production - calls the wrong tool, mangles the output, and cascades into failure. Osmosis, a company out of Y Combinator's Winter 2025 batch, was built around that unglamorous gap between the demo and the deploy.

The premise is contrarian in a market obsessed with size. While much of the industry races to rent the biggest, most general model available, Osmosis argues the opposite: take a smaller open-source model, train it hard on the one task you actually care about using reinforcement learning, and you end up with something faster, cheaper, and more reliable than a generalist giant. The company puts it plainly on its own site - it helps companies "build AI agents that actually work."

The Founders

Two friends who met in middle school

Osmosis was founded in 2024 by Kasey Zhang and Andy Lyu, who met more than a decade ago as kids. Zhang, the CEO, had already co-founded a gaming startup that was acquired, then spent time in venture capital focused on AI and data investments - the vantage point of someone who watched a lot of AI companies stumble on the same problems. Lyu, the CTO, was the youngest tech lead at TikTok, where he ran real-time recommendations and data infrastructure at a scale most engineers never touch.

That split matters. One founder spent years studying why AI products fail to ship; the other spent years building the real-time systems that keep enormous models fed and responsive. Osmosis is the intersection of those two obsessions.

There is a quiet advantage in co-founders who have known each other since they were teenagers. The awkward early conversations - who decides what, how you disagree, when to cut a project - are conversations they have already had, in some form, for years. In a field where technical bets shift monthly and the roadmap has to bend with them, that kind of trust is less romantic than it sounds and more useful than most cap tables.

The Problem

Foundation models are generalists. Your job is specific.

A foundation model is trained to be passably good at nearly everything, which means it is rarely excellent at your narrow, high-stakes task. For a lot of applications that trade-off is invisible. For AI agents - systems that chain together several tool calls to get real work done - it is expensive. Latency stacks up across turns. Tool use gets brittle. One bad step early can topple the whole chain.

Osmosis targets exactly this failure mode. Its platform runs multi-turn tool training over the Model Context Protocol, or MCP, the open standard for how models talk to tools. By replaying something close to production conditions during training - real tools, multiple turns, actual reward signals - the fine-tuned agent learns the behavior it will need on the job, not just the behavior that looks good in a benchmark.

The distinction is subtle but it is the whole game. Most fine-tuning teaches a model to imitate good answers. Reinforcement learning teaches it to earn them - to try, get scored, and adjust. For an agent that has to decide which tool to reach for and when, that difference is the gap between a model that recites the right steps and one that can recover when a step goes sideways. Osmosis is built on the second idea: that reliability is a thing you train, not a thing you hope for.

The Product

Reinforcement learning, minus the headache

The Osmosis platform tries to collapse a messy research workflow into one path. Raw data is processed on-platform - no more, as the company dryly notes, back-and-forth with Jupyter notebooks. From there the system handles compute orchestration, reward modeling, and observability into each training run. It supports real-time reinforcement learning from the start, runs automated hyperparameter sweeps, and can train and maintain hundreds of model versions at once for teams that need many specialized variants.

Tools can be built natively inside the platform or hooked in over MCP, and agentic tools can be synced straight from a GitHub repository to feed tool-use training jobs. When the model is ready, you serve it on single-tenant compute or export it for self-hosting. A detail Osmosis returns to often: you own the weights. Full VPC and on-premises deployment is available for teams that cannot let their data leave the building.

There is also a continuous-improvement loop. Models can retrain automatically on new data - triggered by a time interval or by an evaluation threshold slipping - as often as every hour. A model that keeps learning after launch is a different kind of asset than one frozen the day it shipped.

For teams that need many variants rather than one, Osmosis leans on LoRA adapters - small, cheap-to-train add-ons that specialize a shared base model. The company has written about training thousands of these at once by keeping one base model in memory and routing tokens to the right adapter. Practically, that turns "one custom model" from a project into a menu. A support team could run a different tuned agent per product line without paying to host a separate model for each.

The Difference

Small and sharp versus big and blunt

The clearest way to understand Osmosis is the trade it asks you to make. A general foundation model gives you breadth. A model fine-tuned on Osmosis gives you depth on the task that pays your bills - and, the company argues, wins on the metrics that matter in production.

This lands Osmosis in the fast-moving category of post-training and RL-as-a-service, alongside names like OpenPipe, Predibase, Together AI, and Fireworks AI, plus the fine-tuning features baked into the foundation-model providers themselves. Osmosis's chosen wedge is agents specifically - multi-turn, tool-heavy workloads - and a "forward-deployed" posture where its own engineers sit with customers from dataset curation to model delivery. Forward-deployed is a term borrowed from a different corner of tech; here it doubles as a moat.

The service angle is worth sitting with, because it cuts against the reflex of most software startups to automate the humans out of the loop. Reinforcement learning is finicky. The reward function is where projects live or die, and writing a good one for a real business task is closer to craft than configuration. By putting its own engineers next to the customer during that step, Osmosis is betting that the early market does not want a self-serve dashboard - it wants the training to just work, and it will trade some independence to get there. High-touch is not the enemy of a software business. Sometimes it is the reason the software gets adopted at all.

The Receipts

Open research, in public

For a company selling reinforcement learning, Osmosis does an unusual amount of it in the open. Its blog reads like an applied-RL notebook: a Qwen3-1.7B model fine-tuned for code merging, an open small language model trained specifically for MCP, work on fixing structured JSON outputs, and deeper systems pieces on training thousands of LoRA adapters at once by sharing a base model and routing tokens, and on cutting memory in long-context RL with fused logprobs. It is both marketing and a standing demonstration that the team can do the thing it charges for.

Notice the choice of examples. Code merging, structured JSON, tool use over MCP - these are not flashy consumer demos. They are the plumbing tasks that break silently and cost real money when they do. A model that reliably produces valid JSON is worth more to most engineering teams than one that writes a clever poem, because the JSON is what the next system in the pipeline depends on. Osmosis picking those problems to publish on is itself a statement about who it is building for: the people responsible for the parts of AI that have to work every single time.

The Money

Who is backing it

Osmosis launched publicly in 2025 with a $6.3M seed round co-led by CRV and Audacious Ventures. The angel list reads like a who's-who of people who have shipped infrastructure at scale: Paul Graham of Y Combinator, Guillermo Rauch of Vercel, Misha Laskin of Reflection AI, Erik Bernhardsson of Modal Labs, and Shane Gu of Google DeepMind. When the people who built the tools your tools run on write the checks, it says something about who the pitch is aimed at.

Where It Fits

A bet on ownership

Strip away the RL vocabulary and the underlying wager is about control. The industry default is to rent intelligence from a handful of large providers. Osmosis is selling the alternative - your task, your data, your weights, your deployment - packaged so a normal engineering team can actually use it. Whether that becomes the dominant pattern or a durable niche is the open question. For the teams already tired of watching a rented generalist fumble their specific job, the appeal is not hard to see.

Explore