LangChain open-sourced the harness behind Content Studio, a Slack design agent that has shipped 3,000+ assets and cut its median request cost from about $8 to about 50 cents
AspireAI community · LangChain
On October 27, 2026 LangChain published a detailed account of building Content Studio, a design agent that lives in a Slack thread and returns finished, on-brand assets, and open-sourced the harness. The architecture gives every request its own sandbox where the agent keeps the project between turns, works across a video engine and image tools, and inspects its own renders, with long-running async subagents reporting plan and progress back to the thread and an approval gate before anything ships. The stated method is that models already write competent code but lack the designers' point of view, so the team's fix was encoding design principles, guardrails against common AI design patterns and the reasoning behind each piece of feedback as skills, and expiring the workarounds as models improve. Reported results include more than 3,000 assets across the company including over 80 slide decks, roughly 4.5M views on X, about 2x the likes and 3.7x the bookmarks on the LangChain account, three in four requests now coming from outside marketing, and a median request cost that fell from about $8 in mid-September to about 50 cents. The honest caveat is that these are results from one company's internal agent on its own brand, not a benchmark.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.