Manos Koukoumidis

We're using AI agents to automate everything. Except the engineering of AI agents themselves.

That seemed backwards to us.

What if you could train your own specialized AI agent with just 15 minutes of your time instead of months of AI engineering?

Would you keep renting the same expensive generic models as your competitors, endlessly patching prompts and harnesses until you hit a wall?

Today, we're launching the agent that builds, trains, and continuously improves AI agents. 🚀

Meet FAiE, our Frontier AI Engineer. Describe what your agent should accomplish, and FAiE designs evaluations, diagnoses failures, curates data, selects training strategies, and runs experiments.

15 minutes of your effort. A few hours of automated AI engineering. Your own specialized agent, with its model weights trained for your task.

Already have an agent? Keep your harness. Connect it to @oumi_ai and train the model behind it with SFT, DPO, or GRPO (Private Preview).

The agent that builds your best agent.

Excited to see what you build: platform.oumi.ai
Live demo today at 10 AM PT: oumi.ai/events/fall26

Why do so many agents fail to live up to their promise?
Building AI agents is easy. Making them reliable is hard.

When an agent fails, someone has to figure out why. Was it the prompt? The harness? Or the model underneath? Most teams keep patching prompts and harnesses until they hit a wall.

𝐓𝐡𝐞 𝐛𝐞𝐬𝐭 𝐚𝐠𝐞𝐧𝐭𝐬 𝐰𝐨𝐧’𝐭 𝐠𝐞𝐭 𝐛𝐞𝐭𝐭𝐞𝐫 𝐭𝐡𝐫𝐨𝐮𝐠𝐡 𝐥𝐨𝐧𝐠𝐞𝐫 𝐩𝐫𝐨𝐦𝐩𝐭𝐬, 𝐡𝐮𝐧𝐝𝐫𝐞𝐝𝐬 𝐨𝐟 𝐬𝐩𝐞𝐜𝐢𝐚𝐥-𝐜𝐚𝐬𝐞 𝐩𝐚𝐭𝐜𝐡𝐞𝐬 𝐢𝐧 𝐭𝐡𝐞𝐢𝐫 𝐡𝐚𝐫𝐧𝐞𝐬𝐬, 𝐨𝐫 𝐞𝐯𝐞𝐫-𝐛𝐢𝐠𝐠𝐞𝐫 𝐠𝐞𝐧𝐞𝐫𝐚𝐥-𝐩𝐮𝐫𝐩𝐨𝐬𝐞 𝐦𝐨𝐝𝐞𝐥𝐬.

They'll get better by training the model behind them on their own work: where they failed, what users preferred, and whether they actually completed the task.

That’s what frontier labs and a small number of leading AI companies already do. They don’t stop at prompting. They train the model itself.

Until now, doing that required specialized expertise and significant effort, putting it out of reach for most enterprises.

On November 4, we're showing what changes that in a 35-minute live demo.

𝐖𝐞𝐝 𝐍𝐨𝐯 4, 10:00 𝐀𝐌 𝐏𝐓.

Can't make it live? Register anyway, and we'll send you the recording.

oumi.ai/events/fall26?…

Jeremy Greer at @oumi_ai figured out how Jev was made.

Details on his guesses/analysis + how you can too below:

𝗤𝘄𝗲𝗻 𝟯.𝟱 𝗼𝗿 𝟯.𝟲, based on obscure & famous knowledge comparisons, name probing, and sensitive question responses

𝟯𝟱𝗕 𝗔𝟯𝗕, based on obscure & famous knowledge comparisons, benchmark performance, and serving profile

Tokenizer is a merge of 𝗤𝘄𝗲𝗻 and 𝗴𝗽𝘁-𝗼𝘀𝘀

RLCD is likely just standard 𝘀𝗼𝗳𝘁-𝗹𝗮𝗯𝗲𝗹 𝗰𝗿𝗼𝘀𝘀 𝗲𝗻𝘁𝗿𝗼𝗽𝘆 from synthetic data.

Synthetic data is a mix of 𝘁𝗲𝗮𝗰𝗵𝗲𝗿-𝘁𝗮𝘂𝗴𝗵𝘁 𝗳𝗿𝗼𝗺 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗺𝗼𝗱𝗲𝗹𝘀 and programatic answers to 𝗰𝗼𝗱𝗲-𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗲𝗱 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀.

Serving is likely done with 𝘁𝗿𝗲𝗲 𝗯𝗮𝘁𝗰𝗵𝗶𝗻𝗴 rather than standard vLLM

Based on long document performance, they're likely using a form of 𝘀𝗽𝗮𝗿𝘀𝗲 𝗮𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 or serving on two GPUs and 𝘀𝗽𝗹𝗶𝘁𝘁𝗶𝗻𝗴 𝘁𝗵𝗲 𝗽𝗿𝗲𝗳𝗶𝗹𝗹 between them. Most likely 𝗕𝟮𝟬𝟬𝘀 or above.

All details included in the article! No insider knowledge here so these are my best guesses based on experiments run.

oumi.ai/blog/how-to-bu…

Owning, not renting your AI, is the only way to secure access.

That's why back in March, Cursor built its own coding model.

It was not just about the quality and cost gains; Cursor claimed better quality than GPT/Opus at 1/10 of the cost.

It was also about AI sovereignty.

JUST IN: Nvidia agreed to buy Hugging Face for $12.9B.

That is at 86x the annualized revenue ($150M).
Why so high? This strategic for NVIDIA.