The Agentic Engineer
• Diffusion models are coming for LLMs. Inception Labs dropped Mercury 2, a reasoning model that generates tokens in parallel. 1,000+ tokens/sec on Blackwell GPUs. This changes the math on agentic loops.
• The "agentic IDE" war is heating up. Emdash (YC W26) runs 21 coding agents in parallel. Cloudflare shipped a full agent hosting SDK. Pi Mono is trending hard.
• Hugging Face is standardizing agent skills. Their new skills repo creates a universal format that works across Claude Code, Codex, Gemini CLI, and Cursor. If this catches on, it's the npm of agent capabilities.
Mercury 2: The End of One-Token-at-a-Time
Look, I've been skeptical of "fastest LLM ever" claims. We get one every other week. But Mercury 2 from Inception Labs is doing something genuinely different, and it matters.
Every LLM you use today, GPT, Claude, Gemini, Llama, generates text the same way: one token, then the next, then the next. It's sequential. It's a bottleneck. And when you're running agentic loops that chain dozens of inference calls together, that bottleneck compounds into something painful.
Mercury 2 doesn't do that. It uses diffusion-based decoding, the same family of techniques that made image generation fast. Instead of writing left-to-right like a typewriter, it refines an entire response in parallel, converging over a few steps. Think less "typing" and more "editing a full draft at once."
The numbers: 1,009 tokens/sec on NVIDIA Blackwell GPUs. At $0.25/1M input and $0.75/1M output tokens. Skyvern's CTO said it's "at least twice as fast as GPT-5.2." That's not a marginal improvement. That's a different category.
Here's why this matters for agents specifically. When your agent workflow chains 30 inference calls per task, cutting latency per call from 2 seconds to 0.4 seconds isn't just "faster." It means you can afford more reasoning steps. More tool calls. More retries. The quality ceiling goes up because the cost of thinking goes down.
Zed's co-founder said suggestions "land fast enough to feel like part of your own thinking." That's the bar. Not "fast for an LLM," fast enough that the human doesn't notice the machine.
The catch? Quality is described as "competitive with leading speed-optimized models," not frontier models. This isn't replacing Claude or GPT for complex reasoning. It's carving out a new lane: the model you use for the 80% of agent calls that need to be good-enough and instant.
I think this is the real unlock for production agents. Not smarter models, faster ones. The bottleneck was never intelligence. It was latency compounding across loops. Mercury 2 is the first model that feels like it was built for agents, not adapted for them.
Read more: Inception Labs blog
Emdash: The Agentic IDE
A desktop app that lets you run multiple coding agents in parallel, each isolated in its own git worktree. Supports 21 CLI agents including Claude Code, Qwen Code, Amp, and Codex. Pass Linear/GitHub/Jira tickets directly to agents, review diffs, and merge, all from one UI. Works locally or over SSH. This is what "agentic development" actually looks like: not one agent doing everything, but many doing focused tasks simultaneously.
Cloudflare Agents SDK
Cloudflare quietly shipped one of the most complete agent hosting frameworks I've seen. Each agent is a Durable Object with persistent state, WebSocket support, scheduling, MCP server/client, SQL storage, and React hooks for frontends. Agents hibernate when idle and wake on demand. The killer feature: run millions of them and pay nothing when they're inactive.
Hugging Face Skills
HF released a standardized skill format for coding agents. Each skill is a self-contained folder with a SKILL.md that works across Claude Code, OpenAI Codex, Gemini CLI, and Cursor. Think portable agent capabilities: install a "gradio" skill and your agent knows how to build Gradio apps regardless of which LLM is driving. Could become the package manager for agent knowledge.
Pi Mono: Agent Toolkit Blowing Up
Mario Zechner's monorepo for building AI agents hit 16.3k stars with 3,000 new ones this week. Full stack: unified LLM API, agent runtime with tool calling, coding agent CLI, terminal UI, web UI components, and vLLM pod management. The "shitty coding agent" branding is peak developer marketing.
Diffusion Language Models: A Survey of Methods and Applications
What it is: A comprehensive survey of diffusion-based approaches to language generation, the same technique powering Mercury 2. Covers discrete diffusion, continuous diffusion over embeddings, and hybrid approaches.
Why it matters: Autoregressive generation (one token at a time) has been the only game in town for LLMs. This paper maps out the entire landscape of alternatives. If Mercury 2's approach scales, this survey becomes the roadmap for the next generation of language models.
Key insight: Diffusion models can generate all tokens simultaneously and then iteratively refine them. This parallelism is what gives Mercury 2 its speed advantage. The tradeoff is quality per step vs. number of steps. More refinement steps = better quality but slower. The sweet spot for agent use cases is fewer steps (faster) since most agent calls don't need frontier-level reasoning.
For practitioners: If you're building latency-sensitive agent pipelines, understand this tradeoff. Diffusion LLMs let you dial the quality-speed knob in ways autoregressive models can't. Expect more models in this family by mid-2026.
Moonshine: On-Device STT That Actually Works
Open-weight speech-to-text models that outperform Whisper Large v3 on the Open ASR Leaderboard. Models range from 26MB (IoT/edge) to full-size. Runs on everything: Python, iOS, Android, Raspberry Pi, even wearables. All on-device, no API keys.
Install it:
Two lines to transcribe audio:
What impressed me: the smallest model is 26MB and still usable. The larger models genuinely compete with Whisper Large v3 on accuracy. Latency is noticeably lower because it's optimized for streaming, it processes audio while you're still talking.
The multi-platform story is real too. Same library works on Python, iOS, Android, and C++ for embedded. If you're building voice into an agent pipeline, this replaces the "send audio to an API" step entirely.
Honest take: documentation could be better for advanced use cases, and non-English support is newer and less battle-tested. But for English STT on edge devices? This is the new default.
Weekly star tracker — February 25, 2026.
| Framework | Stars | Notes |
|---|---|---|
| OpenClaw | 227,487 | The lobster reigns. |
| n8n | 176,278 | Workflow automation king. |
| Dify | 130,300 | Full-stack LLM app platform. |
| LangChain | 127,389 | The OG, still growing. |
| AutoGen | 54,822 | MS nudging users to Agent Framework. |
| Flowise | 49,334 | Visual agent builder. |
| LlamaIndex | 47,183 | Pivoting to agentic OCR. |
| CrewAI | 44,586 | Role-playing orchestration. |
| Semantic Kernel | 27,307 | Microsoft's enterprise play. |
| Haystack | 24,310 | Context-engineered pipelines. |
The top of the chart is pulling away from the pack. Microsoft nudging AutoGen to Agent Framework is the move to watch.
"Agentic" stopped being a feature and became an infrastructure category. Cloudflare is building agent hosting primitives. Emdash treats agents like IDE tabs. HF is packaging agent knowledge like npm modules. Mercury 2 optimized specifically for agent loop latency. We crossed a line. Six months ago, "agent" meant a chatbot with tool calls. Now it means persistent processes with state, scheduling, and lifecycle management. The companies building agent infrastructure, not agent demos, are the ones that'll matter in 12 months. The demo era is over. The plumbing era just started.
Mass Time App
The world's largest Catholic church finder. 280,000+ churches across 131 countries. Daily readings with AI reflections, 60+ prayers, and a home screen widget. 100% free, 100% private. Built by Waltsoft Inc.
Download Free →Want to sponsor this newsletter? Get in touch
Like what you read?
Forward this to a friend who's building with agents.
Subscribe to The Agentic Engineer