Do agents actually learn, or are you just getting better at maintaining them? A memory-transfer question

artemgetman · reddit · 2026-10-24

A Redditor asks whether agent improvement is real learning or just humans getting better at updating prompts, skills, and memory. He wants an agent that can take an unfamiliar customer problem, build a small service, and improve over months without per-step workflow design.

The key question is transfer: if an agent learns to question misleading evidence while debugging, does that help it question a claim in a customer interview — and can it recognize when the analogy doesn't apply? During Anthropic's Claude Code hackathon he experimented with persistent memory in his Cortex project: some comparisons improved, others favored the baseline, so he's not convinced more memory is the answer. Fine-tuning weights for frequently used skills might help, but should be benchmarked against a strong memory + skills setup.

His test: give an experienced agent and a fresh copy unfamiliar work — does the veteran do better, need fewer corrections, and require less maintenance? He's looking for anyone who has tested transfer across genuinely different tasks.

Original post →

Do agents actually learn, or are you just getting better at maintaining them? A memory-transfer question

More from coding & agent