AI Daily · 2026-09-14

The main thread today is AI agents moving from internal experiments to production: LangChain published results from its GTM and paid media agents and introduced managed credentials in Connections, decoupling permissions, auditing, and rotation from code. Along with Perplexity bringing its Portable Computer agent to Windows with RTX acceleration, the updates emphasize deploying agents into real workflows with clear boundaries between model judgment, code execution, and human approval. On research, Tencent Hunyuan released Simple Attention Sparsification attention-gating checkpoints, using learnable KV-block ranking and continuous gating for sparse long-context decoding; they require the Qwen3 base model and seer_attn backend. Other updates include Credit Genie using OpenWiki to keep repo documentation automatically fresh, and Simon Willison shipping commit-rewriter and WebP screenshot support in shot-scraper 1.12.

East Asia · First-hand

Tencent Hunyuan

⭐⭐ [Research] tencent/Simple-Attention-Sparsification

Tencent Hunyuan released checkpoints for Simple Attention Sparsification (SAS) on HuggingFace. The method learns to rank and select KV blocks for each query; unlike sparse-attention selectors trained by distilling dense attention scores, SAS adds continuous gates to the selected blocks so the language-modeling loss can optimize context ranking end to end. The three released checkpoints are built on Qwen3-4B, Qwen3-8B, and Qwen3-14B, with 33.0M/33.0M/42.0M gate parameters (64/64/81 MiB). The authors state these are router-only checkpoints, not standalone language models: the Qwen3 backbone was frozen during training and is not included, so inference requires the corresponding Qwen3 base model and the seer_attn backend in their sglang-blocksparse fork. All three checkpoints share the same configuration: 64-token KV blocks, a training Top-K of 31 historical blocks, gate hidden size 128, training on OpenR1-Math-220k with 32,768-token sequences, and a default 2,048-token sparse decode budget, evaluable at 1,024/2,048/4,096 tokens without retraining.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Product Update

⭐⭐ [Product Update] Perplexity Portable Computer Is Now Available on Windows, Powered by NVIDIA RTX

An NVIDIA blog post says Perplexity's Portable Computer is now available on Windows PCs, accelerated by NVIDIA RTX. It describes Portable Computer as a local version of the Perplexity Computer agent that plans and carries out multistep tasks. As local models become more capable, agents can handle more work directly on a PC while keeping sensitive information on the device. It uses local models to analyze data and bring together information.

⭐⭐ [Product Update] Connections: managed credentials and per-caller identity for Managed Deep Agents

LangChain introduced Connections for Managed Deep Agents (v0.7.0+), which stores credentials in the LangSmith workspace rather than in .env files or the build, so they can be rotated or revoked without code changes or redeployment. Each connection is defined along two independent axes: owner (agent or caller) and credential type (static secret or OAuth grant). Agent-owned credentials are shared by every caller, while user-owned credentials resolve per person at run time so the agent can act on the caller's behalf. Connections are created with `mda connections create` and read at run time by slug through `connections.get()`. Managed Deep Agents handles the OAuth round-trip, so projects need no callback route, token store, refresh logic, or consent screen; GitHub ships in the connections catalog alongside 22 other services, and any OAuth provider can be added using your own metadata.

⭐⭐ [Product Update] commit-rewriter 0.1

Simon Willison released commit-rewriter 0.1, a small Python web app for editing git commit messages. He built it while preparing the Datasette security releases, whose initial commits were full of coding agent cruft and references to issue IDs from a private repository, making them unsuitable for publication. It is run with uvx commit-rewriter plus the path to a repository, or without the path if you are already in that directory. On submitting edits, the tool creates a timestamped branch of the current repo state so you can revert if needed, then rewrites every commit from the first edited one through the most recent.

⭐⭐ [Product Update] shot-scraper 1.12

Simon Willison released shot-scraper 1.12, adding WebP screenshot support to his CLI tool for taking website screenshots, recording video demos and scraping sites with JavaScript. WebP screenshots are taken by specifying an output file ending in .webp, with an optional --quality flag; without that flag the WebP file is lossless. Willison says WebP screenshots are in his experience almost always significantly smaller in file size than equivalent JPEG or PNG files. He shipped the feature so he could generate the screenshot for his new commit-rewriter tool.

Other

⭐⭐ [Other] How We Built LangChain’s Paid Media Agent

LangChain describes how it built its Paid Media Agent: to scale from zero to five paid channels in six months, the team built a long-running agent that lives in Slack and runs every Monday, combining ad-platform data with business context to track new products, draft campaigns, add keywords, test variations, and surface proposed experiments for team approval. The post offers three engineering lessons: treat agents like knowledge workers with a sandbox, software, and clear instructions; use models for judgment and code for consistency; and design around the full workflow, including permissions and human approval. On results, paid media went from 0% to 20% of the marketing pipeline in six months, cost per qualified lead fell 30% from June to August while monthly spend rose about 60%, LinkedIn CPL was 40% lower than in January, and bringing analysis and reporting in-house saved about $5K per month; moving calculations into code and removing unnecessary model calls made an early reporting workflow about 40x cheaper and 13x faster, cutting runtime from 18 minutes to 85 seconds. The agent has been open-sourced, and the team will hold a webinar on September 23.

⭐⭐ [Other] How Credit Genie keeps codebase docs fresh with OpenWiki

A LangChain blog case study describes how Credit Genie, a mobile-first financial wellness platform, uses LangChain's open-source repo-documentation agent OpenWiki to keep codebase knowledge current. Previously the team relied on Notion pages, READMEs, and AGENTS.md files that quickly went stale and were hard to find, with knowledge increasingly siloed among individual contributors and coding agents often acting on outdated context. Credit Genie uses OpenWiki to automatically generate and update repo-level documentation from code changes, and aggregates those docs across onboarded repositories into a single searchable portal hosted on GitHub Pages. Automation includes nightly OpenWiki runs, automated PRs, auto-merge workflows, and daily portal rebuilds; engineers use the portal to understand unfamiliar systems, while coding agents are prompted to check the openwiki/ folder before making changes. The team's next step is connecting OpenWiki with its internal knowledge graph for cross-repo awareness of dependencies and change impact.

⭐⭐ [Other] How we built LangChain’s GTM Agent

The LangChain team describes the GTM agent it built internally on Deep Agents: it triggers on new Salesforce leads, first checks whether outreach should happen at all (for example, whether the person just filed a support ticket or a teammate already reached out), gathers context including meeting history, then posts a Slack draft with reasoning and sources for a rep to approve. Non-negotiables included human-in-the-loop approval before anything is sent, checking contact history, drafts that reflect the account's current state, visible key inputs for explainability, and a learning loop from rep edits; every send, edit and cancel action is logged to LangSmith and attached to the underlying trace. Beyond drafting, the agent aggregates account-level signals across web activity, developer ecosystems, product usage and marketing touchpoints to surface deal risks, expansion opportunities and competitive moves. Reported results: lead-to-qualified-opportunity conversion up 250% from December 2025 to March 2026 with 3x more pipeline dollars in the same period, reps increasing follow-up on lower-intent leads by 97% and higher-intent leads by 18%, 40 hours saved per rep per month (1,320 hours across the team), and 50% daily / 86% weekly active usage among sales team members.