Executive Summary
Hermes Agent by NousResearch has emerged as a significant alternative to OpenClaw, with rapid user adoption driven by its customizability and performance on local hardware. Nvidia's Cascade 2 model showed impressive speed (187 tok/s) on consumer GPUs but failed code coherence tests, while testing continues on OpenReasoning-Nemotron 32B. llama.cpp reached 100k GitHub stars, cementing its role in local AI infrastructure. Claude Code continues gaining traction for terminal-based development workflows. Open-source developers are increasingly vocal about sustainable models in an AI-accelerated ecosystem.
Key Events
- Hermes Agent sees surge in adoption, with users praising its extensibility and Pi integration compared to OpenClaw → tweet
- Ollama launches Pi integration:
ollama launch pi --model kimi-k2.5:cloud, enabling customizable local coding agents → tweet - llama.cpp hits 100k GitHub stars, recognized as foundational open-source infrastructure → tweet
- Nvidia Cascade 2 tested on RTX 3090 at 187 tok/s but failed coding coherence tests; OpenReasoning-Nemotron 32B testing in progress → tweet
- Microsoft released new multilingual embedding models, noted as high-quality additions to Hugging Face → tweet
- Claude Code adds functionality to launch and test apps directly from terminal → tweet
- Community concerns raised about Claude Opus quality degradation and Pro plan usability → tweet
- OpenCode Go discussed subscription economics: $10/month plans break even only with volume and provider negotiations → tweet
Analysis
The open-source AI agent ecosystem is fragmenting rapidly. Hermes Agent is capturing mindshare from OpenClaw through community-driven extensibility and transparent development. Local AI advocates like @sudoingX are running rigorous benchmarks that expose real performance gaps between marketing claims and actual capability—Nvidia's MoE architecture delivered speed but not coding coherence. The pattern suggests consumer GPU users need dense reasoning models rather than sparse-speed models for agent work.
Long context implementation is being flagged as the next transparency reckoning—frontier labs claim it, but 2026 may expose how it's actually architected under the hood. Meanwhile, subscription economics for AI coding tools remain tight: margins depend on negotiated inference deals, making higher-tier plans financially risky for early-stage products.
Watch next: Hermes community plugin ecosystem growth, outcomes of Nemotron 32B local testing, Claude Opus quality reports, and whether Stripe responds to fintech encroachment on their checkout-to-banking position.
Tweet Feed
AI Agent Ecosystem & Hermes Agent
@Teknium · 2026-03-30T16:17
Love the community working with us on building Hermes Agent. If you want to too, just ask hermes-agent to make the changes you want to see in Hermes Agents' codebase and ask it to submit a PR! → tweet
@Teknium · 2026-03-30T15:34
We ship fast we do what our people want done! <3 → tweet
@Teknium · 2026-03-30T18:12
New plugin by the creator of Axolotl himself for Hermes Agent! Check it out: → tweet
@sudoingX · 2026-03-30T05:51
wow look 3,000+ heralds now in like 6 days only. if you're new to hermes agent, migrating from openclaw bloat, or just getting started, this is where you come. we share real configs, real help, real people who actually run this stuff daily. → tweet
@sudoingX · 2026-03-30T06:43
yesterday i tested nvidia's cascade 2 on a single RTX 3090. 187 tok/s, fastest model i've benchmarked in the 3B active class. but when i gave it octopus invaders, blank screens every time 5 times. 3B active MoE couldn't hold architectural coherence across thousands of lines of game logic. → tweet
@sudoingX · 2026-03-30T08:44
so far i test 2 nvidia models tested on my 3090. cascade 2: 187 tok/s, blank screens on every coding test... openreasoning 32B dense: 36 tok/s, overthinks everything. 5 million deepseek reasoning traces turned this model into a thinker that forgot how to ship. → tweet
@badlogicgames · 2026-03-30T13:46
still not using subagents, but i'm glad pi's extensibility allows others to go nuts. → tweet
@badlogicgames · 2026-03-30T16:43
since all the frontier labs now offer "long context", 2026 is going be the year when everyone finds out how "long context" is actually implemented under the hood :) → tweet
Local AI & Hardware Benchmarks
@sudoingX · 2026-03-30T11:50
the people telling you a single 3090 can't ship production quality are not wrong about the ceiling. they're wrong about the conclusion... every local AI transition destroys someone's SaaS margin. that's why they fight it. → tweet
@jsuarez · 2026-03-30T17:33
B200 off Vast is slower than a 4090 for small networks out of the box with Puffer 4. But if you want to run a 100M param net on breakout for some reason, it can train that at 800k steps/seconds → tweet
Open Source Milestones
@victormustar · 2026-03-30T15:33
llama.cpp hits 100k stars on github 🙌 Some software leaves a permanent mark on the history of computing and llama.cpp is one of them. Showing why open source is important, let's celebrate it 🎉 → tweet
@victormustar · 2026-03-30T08:42
Surprise drop: new multilingual embedding models by Microsoft - seem quite good :) → tweet
@victormustar · 2026-03-30T09:03
"Mr. Chatterbox is a language model trained entirely from scratch on a corpus of over 28,000 Victorian-era British texts published between 1837 and 1899" Chat with it on Hugging Face ⬇️ → tweet
Developer Tools & Claude Code
@nummanali · 2026-03-30T18:00
Orchestrating Codex from Claude Code is not a bad idea! OpenAI dev team have officially released a plugin to use Codex through Claude Code → tweet
@nummanali · 2026-03-30T17:13
Claude Code computer use is pretty cool. If you want the same for other coding agents, I recommend checking out peekaboo cli from @steipete. Does the exact same but in a CLI package → tweet
@FinansowyUmysl · 2026-03-30T17:25
Claude Code wprowadził nową funkcjonalność. Z poziomu terminalu może uruchomić budowaną aplikację i ją przetestować. To jeszcze mocniej przyśpieszy pracę. [Claude Code introduced new functionality - it can launch and test applications from terminal] → tweet
@LinusEkenstam · 2026-03-30T05:52
Can I see some links of what you've built with Claude code? Blow my mind → tweet
Enterprise & Industry
@RealGeneKim · 2026-03-30T07:37
"With AI, we all got rebooted at the same time." Amy Willard, Global IT Director at John Deere, discussing enterprise AI transformation: workforce readiness, 300 parallel experiments, 92% daily AI engagement. → tweet
@thdxr · 2026-03-30T04:50
people often ask if OpenCode Go will have a higher tier than $10/month. subscription plans need to offer more than spending the same amount on pay per token... on the $10 if we get everything right we break even. → tweet
@levelsio · 2026-03-30T14:23
I think @Stripe needs to add payment cards ASAP to spend your balance... Once the other fintechs also add checkout options there's a real switching opportunity for a lot of people on Stripe → tweet
Commentary on AI & Development
@TheAhmadOsman · 2026-03-30T15:02
Spending time learning graphs and networking theory is one of the highest-ROI investments you can make. It quietly compounds across distributed systems, AI, infrastructure, markets, and even social dynamics. → tweet
@juliarturc · 2026-03-29T19:36
Why so many of us feel career-homeless in tech: Startups full of fraud/grifters, FAANG full of politics, Frontier labs in a race with no morals, Academia full of title collectors... → tweet
@thdxr · 2026-03-29T21:56
we all live in a big ecosystem that all feeds into each other - frontier labs push the bar, opensource labs build off that, inference providers make capex investments, app builders create demand... no one spot is morally superior to the others → tweet