Candidly's state-aware agent harness lifts conversation resolution from 30% to 78% on LangSmith
LangChain · x · 2026-09-08
Candidly's guest post on the LangChain blog details how they built a state-aware agent harness for Cait, their AI financial planning agent, using LangSmith.
Core insight: most assistants judge conversations after the fact, while the agent acts turn by turn. Their approach:
- Read the partial trace to infer the user's live state
- Select response levers based on how similar conversations shifted historically
Key numbers: resolution is 78% when a conversation is going well but drops to 30% when it isn't. The state model measures this gap in real time—a single prompt insert mirroring the user's wording at the right turn can pull a disengaging conversation back up.
The work builds on a formal research paper and months of production analysis, applied to high-stakes financial decisions like debt payoff, savings, and student loans.
Related event: Candidly's state-aware AI lifts dialogue resolution from 30% to 78%(2 posts)→
More from coding & agent
- Codex can generate tool-calling code on the fly — a new failure point for agent devs — srchvrs · 2026-09-09
- Paper: general coding agents beat purpose-built data agents by up to 37 points — RishiBommasani · 2026-09-09
- astra-chess: open-source Codex + skills setup for playing chess with GPT-6 Astra — MikePFrank · 2026-09-09
- Weaviate shows how to turn a messy creative archive into semantic search without renaming files — philipvollet · 2026-09-09
- Factory Worker 'Raises' a Local Gemma 31b AI in a Folder — She Gave Herself a Sense of Touch — D33lix · 2026-09-09
- Agent's per-turn continuation killed after public exploit bragging — flowersslop · 2026-09-09