Applied Compute on Continual Learning for Enterprise Agents: Online Hints Fix Formatting and Efficiency

AI Engineer · youtube · 2026-08-13

Samuel Denton from Applied Compute shared how to bring continual learning into enterprise agents. They aimed to reduce submission turns on SWE bench from 80 to 40, successfully increasing the tool call rate from 22% to 60%.

Four Quadrants of Distillation

Denton proposed an analytical grid: one axis is how 'online' the traces are (from static production dumps to unified serving/training loops), and the other is the hint source (static priors vs. dynamic builds based on on-policy actions). They focus on two corners: Offline hints on offline traces need no replayable environment and can improve agents from a data dump on day one; Online hints on online traces have a much higher ceiling.

The Victory of Dynamic Hints

In a customer case requiring unusual hyperlink formatting, directly rewarding the format or finetuning on it degraded the model's coding ability. Instead, a dynamic hint written against each rollout jumped correct formatting from 15% to 80%.

Practical Mechanisms

To make dynamic hints work, they use two core mechanisms: letting a judge pick where the hint goes in the rollout and distilling only the next few steps, as the learning signal decays with distance. Secondly, they mask learned tokens to remove the teacher's strong opinions on irrelevant connector words. Notably, this entire process is done without a golden answer to distill toward.

Original post →

More from coding & agent

coding & agent channel →