Business-state rewards
Booking and CRM checks score what the agent did.
CRM / Recorded reinforcement learning run
Rewards. Model updates.
Another test.
Qwen3.5–9B completed 48 training episodes and nine model updates in the CRM booking environment, followed by 16 development episodes.
7 / 16before training
13 / 16after nine updates
Same CRM development episodes.
Seven improved · one regressed · eight unchanged.
Recorded model updates
Nine optimizer receipts record changes to model weights.
A saved checkpoint
The updated adapter is retained alongside the final evaluation.
104 episodes.
The full execution retained.
40 baseline episodes · 48 training episodes · 16 final development episodes.
Inspect the evaluation and its limits
The recorded development results were 7/16 before training and 13/16 afterward: seven improved, one regressed and eight were unchanged. These are repeated development cases, not held-out tasks. The final observations exposed the target contact ID, so the result does not establish generalised improvement or isolate the effect of training.
Of the 40 baseline episodes, 24 were training-baseline and 16 development-baseline. Three baseline output-cap stops and one training parse failure remain counted as zero. The record retains 46 tool errors. No failed attempts have been dropped.
Build the learning cycle around your work.
Define what your agent needs to do, the tools it uses, and what a correct outcome looks like.