Voice personalization for agents: better evals beat fancier architectures
linamagr · reddit · 2026-09-18
A practitioner highlights an underrated agent problem: making an agent truly replicate one specific person's writing — punctuation habits, hedge words, sentence rhythm, not just a generic tone. Founders he's spoken with converge on one pattern: architecture matters less than evals, and simpler systems with better voice-level evals consistently beat elaborate ones. He poses open questions on the smallest structural signal capturing someone's style, building the eval loop (human raters vs held-out corpus), avoiding regression when adding users to a shared model, and whether fine-tuning beats prompting + retrieval.
More from coding & agent
- Dev uses Codex to phone a comedy club and book tickets with a gift card — brandon_galang · 2026-09-18
- Devin ran a pest-control lead business with a Ramp card—then showed how bad agent intuition still is — sandylikesfrogs · 2026-09-18
- Sim Search lets agents build a knowledge graph across 1,000+ tool integrations — JafarNajafov · 2026-09-18
- Ethan Mollick rebuilds Umberto Eco's 33,000-book library in 3D with Claude Projects — emollick · 2026-09-18
- Cua and typesafeai launch jev-use dev preview, claiming fast computer use is solved — TianbaoX · 2026-09-18
- Developer demos auto-navigating slides driven by the Jev computer-use model — threepointone · 2026-09-18