Kimmy K2.6 Agent Swarm Launches Web Agency in 40 Minutes
Moonshot AI's Kimmy K2.6 triples agent swarm to 300 sub-agents for 4,000-step tasks, generating 20 custom notary landing pages plus outreach emails in 40 minutes—cheaper than Claude for production agentic workflows.
Scale Complex Workflows with 300-Agent Swarms and Preserve-Thinking Mode
Kimmy K2.6 triples agent swarm capacity from 100 in K2.5 to 300 specialized sub-agents, enabling up to 4,000 coordinated steps for parallel tasks without memory drift. Activate preserve thinking mode to maintain consistent reasoning across multi-turn interactions, preventing degradation in long workflows. In tests, five sub-agents handled a 40-minute task: scraping Google Maps and Canadian Yellow Pages for 20 Greater Toronto notaries with outdated or missing sites, analyzing viability, estimating market size/revenue potential, generating tailored outreach emails, and producing landing page files with previews. Follow-up in 17 minutes applied unique styles, CSS animations, scroll effects, GSAP, and custom AI-generated header images to each—boosting visual appeal despite shared boilerplate structure. Outcomes: Ready-to-send proposals and deployable sites turn local research into a side web agency gig, though uniform templates limit full uniqueness without detailed prompts.
Build Full-Stack Apps with Long-Horizon Coding and Native Vision
Leverage MoonVIT vision encoder (open-source on Hugging Face) for coding-driven UI/UX reasoning, converting prompts or visuals into interactive prototypes with auth, database logging, and effects. For a RAM price comparison site, Kimmy delivered in 12 minutes: dark-themed frontend toggling brands/prices from Amazon, Newegg, Best Buy (scraped via Axios/Cheerio); live refresh button; add-to-compare functionality yielding dynamic tables. Backend used bare Node.js/Express with vanilla JS DOM manipulation—no React—prioritizing functionality over frameworks. Fixes for missing images or features required follow-ups, but token tracking in CLI aids cost monitoring. Claimed 185% throughput on 13-hour engineering tasks holds for production: reliable generalization across front-to-back stacks at lower cost than Claude, requiring Allegretto plan for swarms.
Trade-offs: Strong QoL Gains but Iterative Polish Needed
K2.6 isn't a massive leap from K2.5's frontend strengths—incremental wins like horizontal scaling, vision integration, and open-source components shine for indie builders. Pages risk sameness or CSS breaks without precise instructions; scrapers miss some assets. Still, cheaper token efficiency (no limits burned vs. Claude) and standalone usability make it viable for agentic production, especially swarms automating business dev like local site generation.