Tenkai Daily — May 18, 2026

Model Releases

• Microsoft Fara-7B: Multimodal VL model — Qwen2.5-VL based image-text-to-text model under MIT license, with TGI support. Solid if you need a permissively-licensed multimodal model that won’t ask you to sign your firstborn over to a cloud provider. 📄

Open Source Releases

• code-context-mcp 2.0.4 — MCP server that gives Claude Code local RAG over your repo. Stops your AI from hallucinating functions that haven’t existed since 2019. 🛠️

Research Worth Reading

• DeepSlide: From Artifacts to Presentation Delivery — Looks at the gap between “generating a pretty slide deck” and “actually delivering a presentation well.” Spoiler: pacing and narrative matter more than gradients. 📄
• SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State-Constrained Dispatch — SDOF adds stage constraints to LangChain/LangGraph/CrewAI pipelines so they stop pretending business processes are just graphs with vibes. 🛠️
• SkillSmith: Compiling Agent Skills into Boundary-Guided Runtime Interfaces — Instead of stuffing skills into prompts and hoping, SkillSmith compiles them into proper runtime interfaces with boundaries. Sensible engineering approach to the skills-in-prompts problem. 📄
• Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions — Instruction-tuned models can look fair on the surface while carrying biased representations under the hood. This paper asks whether those internals can leak through—and the answer isn’t comforting. 🔥
• CAX-Agent: A Lightweight Agent Harness for Reliable APDL Automation — LLMs running MAPDL simulations tend to break things. CAX-Agent adds structured execution control, tool encapsulation, and fault recovery so your FEA runs don’t become a comedy of errors. 🛠️
• NOVA: Fundamental Limits of Knowledge Discovery Through AI — Models the “generate, verify, accumulate, retrain” loop as adaptive sampling and asks: can AI actually discover new knowledge, and at what cost? The limits it finds are worth thinking about. 📄

AI Dev Tools

• tech-leads-club/agent-skills — Vetted skill registry for Antigravity, Claude Code, Cursor, and Copilot. Think of it as a curated plugin store that won’t let your agents run off the rails. 🛠️
• Light-Heart-Labs/DreamServer — Self-hosted local AI platform: LLM inference, chat, voice, agents, RAG, image gen — no cloud, no subscriptions. For the “I want it all on my hardware” crowd. 🤖

Today’s Synthesis

There’s a quiet theme running through today’s picks: the industry is slowly admitting that stuffing more capabilities into prompts isn’t engineering—it’s gambling. SkillSmith tackles this head-on by compiling agent skills into proper runtime interfaces with boundaries instead of hoping the model respects a markdown comment. SDOF applies the same discipline to multi-agent orchestration, adding stage constraints so your LangGraph pipeline doesn’t treat a procurement approval like a free-form chat. And the tech-leeds-club/agent-skills registry gives you a curated starting point so you’re not hand-rolling every skill from scratch. The through-line: if you’re building anything beyond a toy agent, stop relying on prompt-level conventions and start enforcing structural constraints—at the skill level, at the orchestration level, and at the registry level. The tools and research are converging on the same answer. 🛠️