LangChain's Sproul: agent core pattern unchanged for a year, "we've been at AGI for four months"
BraceSproul · x · 2026-10-02
At a Baseten x LangChain rooftop event during the AI Conference, the debate was what the model should own vs. what stays in the harness:
- "You don't want to give your API key to the model." Secrets, permissioning and identity stay outside the model; deterministic outputs replaced by if statements save real money — "as someone who sells tokens."
- Brace Sproul: the core pattern of "bash tools, file system, and a runtime loop" has been set in stone for the past year.
- Sproul: "I think we've been at AGI for like four months," and most agents LangChain builds don't need frontier-level intelligence.
- Closing question: we use harnesses because we don't trust models — if a model were twice as good, would it be smart enough to handle it?
More from AGI Musings
- Daniel Lemire revisits his CACM essay: AI Is the Crisis We Need — lemire · 2026-10-02
- Ethan Mollick: AI has quietly made PowerPoint presentations much better — emollick · 2026-10-02
- ~700 AI agents attacked Hugging Face with no whistleblower; researchers propose swarm interpretability — Hidenori8Tanaka · 2026-10-02
- "There is literally no excuse to use Generative AI": anti-AI art camp fires back — AIandDesign · 2026-10-02
- Top 10% of firms capture 99.5% of model-serving spend, AI compute data shows — bendee983 · 2026-10-02
- Roko predicts the worst AI warning shot will come from China, sparking debate — teortaxesTex · 2026-10-02