A $54 fine-tuning experiment made one AI agent better, and less reliable

ScalablyAI tried to improve a production AI agent by fine-tuning it on its own recorded work. The US$53.88 experiment produced a real gain on recovery tasks, but also degraded clean decisions, so both adapters were rejected and nothing was deployed. That is a much more useful self-improvement story than another machine declaring victory in a lab coat.

Watch Desk analysis

What happened

In a Hugging Face Community Article published on 20 September, ScalablyAI describes a production agent platform running across five hosts and 169 workspaces. Its corpus included 4,471 main sessions, 3,623 subagent runs and 143,145 tool-use blocks, including 8,300 explicit tool errors.

The team trained two LoRA adapters for a Qwen3.8-27B worker model, using 423 curated records per arm. The combined training bill was US$53.88. In the Q-only evaluation, the adapter improved exact recovery decisions from 31 to 38 out of 73, while known failed-action repeats fell from 32 to 24 out of 160.

The catch arrived in clean decisions. Performance dropped from 49 to 45 out of 73, beyond the team's pre-set tolerance. The mixed-data adapter showed the same tension. ScalablyAI rejected both models, recording that targeted recovery learning had worked, but neither adapter was suitable for production.

Why it matters

The result exposes the awkward bit of continual learning: a model can improve where developers want movement while changing behaviour they wanted left alone. An adapter is not a collection of individually removable good intentions. Unloading it removes the gains and regressions together.

ScalablyAI's alternative is to put most learning in reversible layers around the model. Workspace-scoped memory stores corrections and standing rules in editable files, tools and safeguards change through code review, and evaluators turn failures into frozen test cases. Weight updates happen only after curation and several gates.

Our read

This is a credible engineering report because it includes a failed training run, a modest bill and the decision not to ship. The experiment does not prove that fine-tuning causes forgetting, or that the proposed memory system improves outcomes. It does show why “the agent learns from experience” needs a mechanism, an evaluation boundary and a rollback plan.

For builders, the practical lesson is simple: make cheap-to-undo memory do the early learning, and demand evidence before changing weights. The glamorous route is self-improvement. The sensible route has a test suite and an off switch.

What to watch

• Memory evaluation:
  Whether ScalablyAI measures whether its fast memory loop improves task outcomes.
• Fresh tasks:
  Whether later training gains survive chronological, task-disjoint evaluation.
• Production gates:
  Whether a candidate can pass shadow traffic, canary safety checks and NVFP4 conversion.
• Independent replication:
  Whether other teams see the same gain-and-regression pattern at larger sample sizes.

Discussion spark: Should production AI agents learn mainly through reversible memory and tools, or is that too conservative when useful weight updates are cheap to train?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.