Krea 2 Large, Hermes Agent Use Cases & Coding Agent Costs
Welcome back. A lot happened today — so let's get into it.
Krea.ai just dropped Krea 2 Large, and this is a big one. It's their first foundation model built from scratch — designed specifically around aesthetic diversity and stylistic control. What does that mean in practice? Text-to-image that actually handles vague prompts without falling apart. Style transfer that can pull stylistic components from up to four reference images simultaneously. And a moodboard feature that goes beyond style — it's analyzing concepts and ideas from your references. Krea cofounder Diego Rodriguez says the model really shines on retro art and — I love this — dank memes. He says that signals "something special." Make of that what you will.
Now, how does it actually stack up? Ben over at Contra Labs ran a proper benchmark. Krea 2 Large came in second for style transfer fidelity — which sounds great — but here's the catch. The gap between Krea and the top dog, GPT Image 2, is only 0.14 points. Meanwhile, Krea's gap to the broader field is more than four times larger. So it's close at the top but trails significantly below. Head-to-head win rate sits at 39.5% — fourth place. The model also tends to be conservative — treating your reference images more like hard constraints than loose inspiration. That could be a feature or a bug, depending on what you need.
Switching gears — the Hermes Agent ecosystem is on a tear right now, and the use cases people are building are genuinely wild.
First up — @sudoingX loaded Qwen 3.6 27B dense, Q4 quantized, onto a single RTX 3090. Twenty-one gigs of VRAM. Running at about 41 tokens per second with a 262,000 token context window. Then he used the Hermes Agent harness to one-shot a full game — Octopus Invaders. Eleven files. 2,411 lines of code. The agent ran an autonomous browser loop — debugging, verifying, reloading — and the game was playable on first load. Total wall time: sixteen minutes and forty-one seconds. For context, the previous Qwen 3.5 27B needed a manual scope fix to get there. The 27B didn't.
Same person — @sudoingX — also ran Step-3.5-Flash-REAP-121B — that's 11 billion parameters active per token — on a DGX Spark using unified memory, again through Hermes Agent. He kicked off the agent while stuck in Bangkok traffic. The agent installed dependencies, ran eight tests across three suites — autonomously. All passed by the time he got home.
Then there's this story from @KSimback. They tasked Hermes Agent with analyzing 558 posts from a specific X account — past 60 days. The agent identified 22 stock picks. 21 of them were up in the last 30 days. Average one-month performance: plus 63%. Best performer: up 248%. The agent then expanded to a larger corpus and identified the top photonics stock pickers on the platform — including an account called @crux_capital_.
And Nous Research themselves — the team behind Hermes — integrated something called cua-driver from @trycua into Hermes Agent. This enables background Mac computer use — clicking, typing, scrolling — with any model, and crucially, no screen takeover. The demo showed the agent summarizing Stripe emails and searching an inbox autonomously.
On the tooling side — @Teknium fixed table rendering in the Hermes Agent TUI. Apparently it's now "perfect" — his word. And a contributor named @elok_lam shipped a major upgrade to the Hermes Kanban dashboard — batch select, multi-card drag and drop, bulk actions, full-text card search, and better keyboard accessibility. Most of it got merged.
Now — here's a story that tells you a lot about where enterprise AI adoption actually is right now. Amazon has apparently rolled out an internal tool called MeshClaw — based on OpenClaw — with a public internal dashboard that tracks token usage across teams. The catch? Developers are being measured against an 80% AI usage KPI. So workers are running agents on trivial, meaningless tasks just to hit the number — burning tokens uselessly to avoid triggering layoffs. The incentive structure is completely broken. You're measuring token consumption, not outcomes.
Speaking of OpenClaw — @everestchris6 built something genuinely clever with it. A fully automated lead-gen pipeline for contractors. The agent scrapes new contractor licenses, phone-matches them to Google Business profiles for verification, grabs van photos, uses AI to render a personalized vehicle wrap mockup, then mails a physical postcard via a direct mail API. From license filing to mailbox — fully automated. That's the kind of concrete, real-world application that actually makes sense.
One more OpenClaw note — @FaztTech pointed out that the OpenClaw GitHub repo has surpassed Linux in star count — but Linux is embedded in basically everything on the planet and OpenClaw has almost no production dependencies. Stars are not adoption. Worth remembering.
Now let's talk about the coding agent space — because there are a few stories here that deserve real attention.
First — a security issue. @DarkNavyOrg exposed a flaw in Cursor's "Auto-Run in Sandbox" mode. The vulnerability: agents fetch content from remote URLs, which opens the door to prompt injection attacks. Malicious content from those URLs can execute unauthorized commands — outside the sandbox — without asking the user. That's a significant trust boundary being crossed.
On a more constructive note — Google quietly launched 13 official Agent Skills on GitHub, under the repo google/skills. These are standardized skills designed to work across Claude Code, Cursor, Gemini CLI, GitHub Copilot, and a platform called Antigravity. The goal is interoperability — one skill definition that any of those agents can pick up and run. That's a meaningful infrastructure move.
There's also a very practical cost optimization story going around. @DeRonin_ detailed a model routing system that slashed their monthly AI coding bill from $4,200 down to $312. How? By being surgical about which model touches which task. Claude Opus 4.6 only handles the top 10% — architecture decisions. Kimi 2.6 at $0.50 per million tokens takes implementation, debugging, refactoring. Haiku 4.5 handles lint, formatting, single-line edits. And Qwen 3, running locally via Ollama, handles boilerplate and autocomplete. On top of that — prompt caching cut input tokens by 90%. The math works. You don't need your most expensive model writing your import statements.
And finally — TrainLoop and Mercor published research on Anthropic Sonnet 4.6 running as an APEX-Agent on production logistics tasks. The headline number: success rate jumped from 23.7% on an MCP harness — up to 42% on a Claude Code harness. That's 80% fewer failures. And 40% lower cost per run. The harness matters more than people realize.
That's your AI digest for 13 May 2026.