Read time: 3 mins
Today's collection of AI updates highlights several releases and tutorials intended to make large context windows and agent workflows more practical and efficient.
Agent SKILLS: why concise context matters
A featured tutorial explains that agents with large context windows (around 200k tokens) often accumulate tool definitions, reference docs and brand guides that bury the few hundred tokens of actual instructions they need (the piece cites roughly 400 tokens). When those crucial instructions are buried, agents ignore them and fail in production. The tutorial describes the anatomy of Agent Skills, emphasizes that the first two lines of SKILL.md are often the most important, and shows how an LLM can route queries to the correct skill without embeddings or a separate retrieval layer. It lists five parts that make skills effective and encourages shipping a first skill for a weekly workflow.
SubQ: fully subquadratic sparse attention
SubQ announced a model built on a fully subquadratic sparse attention (SSA) architecture. Instead of comparing every token to every other token (quadratic cost), SubQ focuses only on token relationships that matter, yielding near‑linear compute growth with context length.
Key claims and numbers:
- Context scale: the model can process up to 12 million tokens of context in a single pass.
- Benchmarks: SubQ 1M‑Preview scored 95% on RULER 128K and 81.8 on SWE‑Bench Verified, placing it on par with or ahead of Opus 4.6 and Deepseek V4 Pro for long‑context accuracy and code tasks.
- Speed and efficiency: the sparse attention reportedly runs 52× faster than FlashAttention at 1M tokens while using 63% less compute.
- Cost claim: the SubQ team states that at scale their cost is under 5% of Opus 4.7; the article gives an illustrative comparison of RULER 128K runs costing $8 on SubQ vs. about $2,600 on Opus, but notes that this figure should be treated cautiously.
The SubQ API, a coding agent called SubQ Code, and a long‑context search named SubQ Search entered private beta today; early access is available via subq.ai.
TinyFish: free Web Search and Fetch for developers and agents
TinyFish made its Web Search and Fetch endpoints permanently free with generous rate limits and no credit card required. Search returns structured JSON intended for agents; Fetch renders any URL in a real browser (full JavaScript, SPAs, anti‑bot handling), strips unnecessary content and returns clean markdown. TinyFish runs these services on its own Chromium fleet, which the company says allows both speed and free access.
The endpoints work with various agent frameworks including Claude Code, OpenClaw, Hermes Agent, Cursor and Codex, and are available via API, MCP, Python and TypeScript SDKs, CLI and Skills — all with one API key and no credit card.
Stop RAG‑ing, start grepping with Scout (Agno)
The article argues that Retrieval‑Augmented Generation (RAG) often produces stale indexes and poorly chunked data. Agno's open‑source Scout takes a different approach by “navigating” live information sources at query time. Scout connects to Slack, Google Drive, Linear, MCP servers and others, walking each source's native API to assemble answers with real citations. As it operates, Scout builds a wiki and CRM (Postgres‑backed), even creating new database tables on demand. The project is ready to fork and customize.
Notable design choices:
- Context providers: Scout wraps each source behind a thin sub‑agent, so the main agent sees a single unified interface (e.g., query_slack) rather than many API endpoints.
- Navigation over search: by querying live APIs, newly posted Slack messages become immediately addressable and citations point to openable paths.
- Self‑building CRM/wiki: the agent constructs structured knowledge as it interacts.
Scout ships with Docker Compose, connects to Agno's AgentOS for multi‑user and scheduled tasks, and offers Slack integration with full thread history.
Quick bites and tooling
- Codex received animated “pet” status overlays that visualize whether the agent is working, waiting, or needs review.
- Anthropic released ten agent templates tailored for finance workflows (pitchbooks, KYC screening, month‑end close, valuation checks, etc.) that plug into Cowork and Claude Code or run as Managed Agents; these templates link to data sources like Moody's, Third Bridge and S&P Capital IQ. Claude also now runs inside Excel, PowerPoint, Word and Outlook with context that follows across apps.
- Hermes Agents can become video editors via the HyperFrames Agent Skill by HeyGen. Installing the skill (command shown as $ hermes skills install hyperframes) lets an agent treat HTML as the canonical source for video, taking inputs such as an X post, a PDF or a GitHub repo to script, animate with GSAP, add captions over TTS narration and render a finished MP4 end‑to‑end.
Designlang, Interact AI and fireworks‑tech‑graph
- Designlang: extracts a website's full design system from the live DOM with one command and emits 17+ files (DTCG tokens, Tailwind config, shadcn theme, Figma variables, motion tokens, typed component anatomy, brand voice, page‑intent labels and prompt packs for various tools).
- Interact AI: replaces static sites with adaptive, conversational pages that recompute per visitor in real time so different roles (founder, CISO) see different content generated from the same data.
- fireworks‑tech‑graph: turns plain system descriptions into polished SVG and PNG technical diagrams with multiple visual styles and built‑in knowledge of common AI/agent patterns (RAG pipelines, Mem0 memory layers, multi‑agent flows).
Bottom line
These announcements and tools tackle two related problems: reducing the compute and cost of truly long contexts (SubQ's sparse attention approach) and improving how agents access fresh, relevant context at query time (TinyFish, Scout). At the same time, new skills and integrations make it easier to automate multimedia and enterprise workflows with agents.
