Oumi expands FAiE to build agents and train their models

Oumi is extending its FAiE engineering agent from custom model development to the agents those models power. Its November 4 Fall Launch brings agent construction, evaluation, training data and model improvement into a connected workflow. Teams can also connect an existing agent runtime through Bring Your Own Harness, or BYOH.

The practical attraction is testing a model in the environment where it actually makes decisions. The important qualifications are equally concrete. BYOH needs an integration endpoint, usage is metered across several services, and the GRPO reinforcement learning feature remains in private preview.

What FAiE adds to the agent workflow

The current platform description covers conversations containing user messages, tool calls and tool results. Tools can be simulated or connected to real endpoints. Evaluation can inspect the trajectory, including which tool the agent chose, its arguments and whether the task finished. Failure analysis groups recurring problems so a team can generate targeted training data. Training options include supervised fine tuning, on-policy distillation and direct preference optimization using preferred and rejected response pairs.

That gives builders a more useful question than whether the final answer sounds plausible. Did the model select the right operation, supply the required information and handle the next turn correctly? Those checks still need a task definition and a trustworthy evaluation standard.

Oumi already offered much of the underlying model workflow. Its August summer launch described synthetic data, LoRA, on-policy distillation, production deployment and monitoring. The November story is the expansion toward complete agents and their existing runtimes, rather than the first appearance of every training or deployment feature.

Bring Your Own Harness still needs an adapter

Oumi’s harness overview says the customer keeps prompts, tools, memory and state in its own runtime. Oumi sends conversation inputs to that runtime and supplies a scoped connection to the selected model. The harness returns its messages for evaluation or dataset creation. Keeping the runtime does not mean conversation data never passes through Oumi.

The registration guide requires project Editor access and an endpoint reachable over the public internet using HTTPS on port 443. An existing agent therefore needs to expose the supported interface. This is a compatibility requirement to plan for before committing to a trial.

The harness protocol defines submission, status, output collection and cancellation routes. Each HTTP request must return within 30 seconds, while asynchronous processing can continue for up to 15 minutes per turn. The integration also has to preserve session state and avoid duplicating work when requests are retried.

There are model restrictions too. Custom Oumi models support harness evaluation but not harness synthesis. API hosted models support both. Harness runs use the customer’s tools, so results depend on that actual tool implementation rather than Oumi’s simulated versions.

GRPO is a private preview with specific requirements

The launch and platform pages explicitly label group relative policy optimization, or GRPO, as private preview. Public documentation should not be mistaken for general availability, and the reviewed pages do not establish a general release date or a public preview price.

Oumi’s reinforcement fine tuning guide describes rewards supplied by a standard LLM judge with scoring labels. The documented setup requires a paid plan, which does not guarantee preview access, a supported base model, separate training and validation data, and a configured provider key for the judge. Rubric, classification and extraction evaluators cannot be used as its reward evaluator.

This matters when interpreting the promise of learning from task success. The documented setup depends on a judge that can recognize success reliably. Oumi recommends testing that judge against human labels and held out examples to catch reward gaming. It describes multiphase training as coming soon, so a fully available chained training pipeline should not be assumed.

The subscription is only part of the bill

The pricing page lists Pro at $25 per month, including $25 in platform credits, with usage charged beyond that allowance. Free access offers up to $50 in credits for three months. Its FAQ specifies $50 for a corporate email address and $25 for a personal address. Enterprise pricing is custom.

Agent tokens, evaluation, synthesis, storage, training and inference have separate rates. On-policy distillation currently runs on eight GPUs, with prices quoted per GPU hour. At the listed $2.90 A100 rate, eight A100s would cost $23.20 for one hour of training compute. Hardware is subject to availability, and other metered work adds to that calculated example.

Pro lists one model weight download per month. Teams planning portability should check both the plan and the chosen base model’s terms. Our guide to Liquid AI’s LFM3 family illustrates why downloadable weights and unrestricted commercial licensing are separate questions. This is a licensing example, not a claim that Oumi supports LFM3.

Keep a human acceptance test

The FAiE guide says cost incurring or consequential actions normally require approval, with optional automatic approval. It also lists boundaries, including file imports, billing and role management. FAiE can coordinate engineering work, but the team still owns its acceptance criteria.

For a first trial, keep the tool environment fixed and compare the original model with the trained candidate on unseen tasks. Track task completion, unwanted actions, latency and the full bill. That is our suggested evaluation plan, not a ByteForward benchmark. The value of this release will depend on measurable improvement in a specific workflow, not the number of steps the platform can automate.