Fine-tuning
Continuing a model’s training on domain examples to shift behaviour — terminology, format, tone — into the weights. Powerful when in-context approaches plateau, at the price of maintenance per model version.
What it is genuinely good at
Fine-tuning moves behaviour into the weights: a house format followed without being asked, a domain vocabulary used correctly, a tone held consistently, or a small model taught to do one narrow task at a fraction of the cost of a large one. It is a behaviour tool far more than a knowledge tool.
What it is usually the wrong answer for
Facts that change. Baking a current price list or policy into weights produces a model that is confidently out of date the moment either changes, with no way to correct it short of retraining. Knowledge that moves belongs in retrieval, where it can be edited.
The maintenance nobody budgets for
A fine-tune is attached to a base model version. When the base is deprecated — and it will be — the work must be redone against its successor. That recurring cost is why most teams should exhaust prompting, skills and retrieval first, then fine-tune only where a measured plateau justifies it.
In AGNT
Most AGNT teams get further with skills and memory first; fine-tuned checkpoints plug in via local or hosted providers when justified.
Fine-tuning — common questions
How much data does fine-tuning need?
Fewer examples than people expect for style and format work — often hundreds rather than millions — but they must be consistent. Contradictory examples teach contradiction.
Should I fine-tune before trying skills and RAG?
Almost never. Prompting, skills and retrieval are faster to iterate and easier to correct; fine-tuning is worth it once those have measurably run out of room.
Can AGNT use a fine-tuned model?
Yes. Point it at a fine-tuned checkpoint through a hosted provider or run it locally — to the runtime it is just another provider endpoint.