Frontier Notes / Daily Signal Report


Issue —  · 2026-06-29  · 8 signals


Today


This week, Mixture of Agents (MoA) in Hermes Agent lets users combine multiple open-source models to surpass closed-source frontier models like GPT-5.5 and Opus 4.8, while Stanford's STORM method packages multi-perspective AI research into a free Claude skill that outperforms deep research.

Editor's Notes


The week's videos highlight a shift toward democratizing advanced AI capabilities: from running entire businesses inside ChatGPT to deploying small on-device models that match frontier performance. Key themes include agent orchestration (MoA, STORM, NanoClaw), practical tooling for non-technical users (Claude Design 2.0, AI businesses in ChatGPT), and cost/security optimization via SLMs. Together, they signal a move away from reliance on single monolithic models toward composable, multi-model systems.

Key Takeaways

  1. Use Mixture of Agents (MoA) in Hermes Agent to combine multiple open-source models for frontier-level intelligence without waiting for unreleased closed-source models.
  2. Apply Stanford's STORM method as a free Claude skill to generate multi-sourced, peer-reviewed research briefings faster and cheaper than deep research.
  3. Run five productized services (founder brand system, sales page, branded reports, SOP decks, investor decks) entirely inside ChatGPT to replace expensive agency work.
  4. Adopt the 'prototype big, deploy small' framework with tools like Arize Phoenix to replace cloud-based frontier models with on-device SLMs for cost, latency, and security gains.
  5. Leverage Claude Design 2.0's new canvas editing, design system import, and Zapier MCP connectors to build functional apps with backends like Supabase without coding.
  6. Implement NanoClaw's blueprint for autonomous work agents: prioritize personal agents with memory, containerization, and human-in-the-loop approval for safe enterprise deployment.
  7. Use the 'writing great skills' checklist (trigger, structure, steering, pruning) from Matt PCO's repo to build robust agent skills.
[01] llm 4 signals

Hermes Agent + Mixture of Agents is insane…

Mixture of Agents (MoA) is a new feature in Hermes Agent that consults multiple AI models from different providers in parallel, then uses a single powerful aggregator model to produce a final response, allowing users to surpass the capabilities of top closed-source models like GPT-5.5 and Opus 4.8. The video demonstrates setting up Hermes Agent on a Hostinger VPS, configuring an MoA preset with four reference models and an aggregator, and using it to autonomously build and deploy a full-stack 3D game. The key benefit is achieving frontier-level intelligence without relying on unreleased frontier models from OpenAI and Anthropic, though it comes with higher cost and latency.

[llm] [agents] [mixture-of-agents] [hermes-agent] [open-source] [vps]


GLM-5.2 vs Claude Opus 4.8 – Does GLM REALLY Beat Claude?

GLM 5.2, an open-weights model, is compared head-to-head against Claude Opus 4.8 across several challenging tasks: a 3D skydiving simulator, a Windows XP AI chat app, a 3D-printable V8 engine housing for an N20 motor, and a time-traveling city block scene. Both models performed similarly on the skydiving simulator and the chat app (both eventually working), but Claude Opus 4.8 had a clear advantage in 3D asset design and the STL motor housing task. The city block scene was impressive from both, with Claude's 1945 era being particularly detailed.

[llm] [comparison] [coding] [3d-modeling] [open-weights] [agents]


The Blueprint for Autonomous Work Agents | Gavriel Cohen, NanoClaw

Gavriel Cohen, creator of NanoClaw, outlines a blueprint for autonomous work agents, emphasizing personal agents over team-managed agent factories for enterprise adoption. The Singapore Minister of Foreign Affairs' use of NanoClaw as a second brain with a memory system highlights the killer use case. Security isolation, containerization, and human-in-the-loop approval are critical for safe enterprise deployment.

[llm] [agents] [security] [enterprise] [open-source] [memory]


Frontier results, on device - RL Nabors, Arize

Rachel Lee Neighbors argues that developers can drastically cut costs, improve latency, and enhance security by replacing cloud-based frontier models with smaller, on-device language models (SLMs) for many tasks. She presents a four-step framework—prototype big, deploy small—and demonstrates using Arize's open-source Phoenix tool to evaluate models like Llama 3.2, Qwen, and Gemma, showing that with careful prompt engineering and post-processing, a small model can match or exceed a frontier model's performance on a specific summarization task.

[llm] [slm] [on-device] [cost-optimization] [evaluation] [prompt-engineering]

[02] ai-business 1 signal

5 AI Businesses You Can Run Inside ChatGPT (This Is Dangerous)

Five productized services can be run entirely inside ChatGPT, generating deliverables that agencies used to charge high prices for. A single person can now produce a founder brand system, a sales page, branded reports, SOP training decks, and investor decks in days rather than weeks. The key insight is that these are not new ideas, but existing overcharged services that can be delivered more efficiently with AI tools.

[ai-business] [agency-services] [productized-services] [chatgpt] [gamma] [image-2-0]

[03] agent-skills 1 signal

Building Great Agent Skills: The Missing Manual

The talk presents a checklist for writing great agent skills, covering trigger (user vs model invoked), structure (steps, reference, and making skill.md small), steering (using leading words and hiding future steps for leg work), and pruning (removing duplication, sediment, and no-ops). It introduces the 'writing great skills' skill from Matt PCO's skills repo as a hands-on tool to apply the framework.

[agent-skills] [skill-design] [llm-agents] [prompt-engineering] [best-practices]

[04] storm 1 signal

Stanford's Method Turns Claude Into a PHD Level Research Team

Stanford's STORM research method, which uses multiple expert perspectives to produce more organized articles, has been packaged into a free Claude skill. The skill runs five agent perspectives (practitioner, academic, skeptic, economist, historian), maps contradictions, synthesizes findings, and peer-reviews citations, generating a verified HTML briefing. Compared to Claude's deep research feature, STORM is faster, cheaper, and produces more actionable, multi-sourced reports.

[storm] [multi-agent] [research] [claude] [skills] [subagents]

[05] anthropic 1 signal

Anthropic Just Dropped Claude Design 2.0 (Endless Capabilities)

Anthropic revamped Claude Design with a new interface, unified usage limits across Claude products, direct canvas editing, a design system import from Claude Code, and connectors including Zapier MCP. The platform now supports sending designs to Claude Code and other tools, enabling non-technical users to build functional apps with backends like Supabase.

[anthropic] [claude-design] [design-tools] [ai-prototyping] [connectors] [supabase]

Frontier Notes · Generated Jun 29, 2026 · 8 of 8 signals
You received this as a Frontier Notes recipient.