Own your agent's decisions. Decision models pick an answer instead of writing one. Think Jev, Clef or OpenAI's Decisions API. Overmind trains your agent on those decisions. The weights are yours.

Overmind’s first week as the open-source training platform for agents. Train the model behind your agent. Own the weights. Onwards.

MODEL TRAINING Fine-tune an open model on your own data. Overmind handles training and deployment. You own the weights, with the choice to host on Overmind or run it yourself. Start free: console.overmindlab.ai Star the repo: github.com/overmind-core/ove…

AGENT TESTING & EVALS Overmind scores prompts and models with evals built from real tasks in your production data. Compare results with your current model, inspect failures, and catch regressions before deployment.

DATA WORKSHOP Datasets built from your traces. Our agent cleans, filters, and prepares examples for training and evaluation, capturing the knowledge specific to your product.

AGENT OBSERVABILITY Connect your repository and production traces to see how your agent makes decisions. Overmind inspects the traces, the tool calls, and the context behind it.

Overmind turns anyone into an AI lab. • Builds a context graph from code and traces • Curates training and eval datasets • Evaluates prompts and models • Trains smaller, specialized models github.com/overmind-core/ove…

Where Jev stayed ahead: TabFact (200-row pilot). Jev 92%, best fine-tune 81%. Sub-4B models trailed Jev on every pilot task. Rent the generalist. Train the specialist once the labels exist.

Jev's confidence is its strongest feature, so this run measured it on all 8,556 answers. When 90%+ sure on yes/no questions, Jev was right 96.5% of the time. Calibrated. • 77 banking intents: claimed 98%, right 92%. • 5 sentiment levels: claimed 95%, right 71%.

Three public tasks, full eval splits, 8,556 examples. Jev 1.13 zero-shot with label descriptions. • Banking77: 89.7% vs 80.4% • SST-5: 59.6% vs 56.7% • BoolQ: 91.3% vs 90.0% Lower latency and lower cost per call on all three.

Jev vs Clef settled one thing: a model anyone can rent is not a moat. The labels are. A 4B model trained on 10k labelled banking queries made 47% fewer errors than Jev. One LoRA run, no hyperparameter search, under an hour on Overmind.