Frontier Notes — Jun 28, 2026

Today

Hermes, the successor to OpenClaw, dramatically simplifies AI agent setup with pre-installed skills, plugins, and multi-provider support, making it accessible for non-technical users.

Editor's Notes

This week's videos highlight a strong push towards practical, cost-effective AI deployment: from token-saving strategies for agents to new open-source models like Ornith 1.0 that compete with frontier models. Meanwhile, the ecosystem around OpenClaw evolves with both a physical terminal and a beginner-friendly successor, reflecting a maturing landscape for both developers and end-users. The fictional GPT-5.6 announcement underscores real pressures on pricing and performance from Chinese labs and hardware innovations.

Key Takeaways

Implement token cost reduction strategies: cache system prompts, route queries by difficulty, offload tool results, cap loops, and trim history.
Use Hermes for easy agent setup if non-technical; consider OpenClaw for more control and customization.
For self-improving systems, adopt a bucketed approval strategy and a bias-to-action mindset to balance automation and risk.
Evaluate Ornith 1.0 (especially the 35B MoE variant) for local coding tasks as an open-weight alternative to restricted frontier models.
Monitor pricing trends: Chinese labs like Deepseek and Qwen offer competitive performance at lower cost, and hardware like Cerebras enables high throughput.

[01] llm 2 signals

AI ENGINEER · 1D AGO

OpenClaw in Your Hand: Building a Physical AI Terminal - Lech Kalinowski, Callstack

Lech Kalinowski built a physical AI-native terminal that combines an OLED and e-paper display to serve as both a remote control for OpenClaw on an NVIDIA DGX Spark and a text-based RPG gaming device. The system uses an ESP32 microcontroller, a custom backend with a 120B parameter open-source LLM served via TensorRT, and implements a dual-display, static-buffer rendering approach to keep firmware small and efficient. Key technical challenges included power supply regulation, I2C implementation, and encoder noise, and the project has led to a provisional patent for quiet, distraction-free AI terminals.

[llm] [hardware] [terminal] [agents] [openclaw] [rpg]

→ Watch on YouTube · → Full summary

AUSTIN MARCHESE · 1D AGO

How to Build A Self-Improving System with Claude Code

Building a self-improving system with Claude Code requires a five-step framework: setting up a knowledge base and skills, ingesting historical data, creating continuous data pipelines, designing improvement loops with a bucketed approval strategy, and adopting a bias-to-action mindset. The system augments rather than fully automates, letting AI handle low-risk changes while humans decide on higher-stakes improvements.

[llm] [agents] [local-models] [self-improving-systems] [workflow-automation] [knowledge-base]

→ Watch on YouTube · → Full summary

[02] token-costs 1 signal

AI ENGINEER · 1D AGO

Your Agent Is Wasting Tokens and You Don't Know It - Erik Hanchett, AWS

Erik Hanchett from AWS presents five practical strategies to reduce token costs when building and deploying AI agents: caching system prompts, routing queries by difficulty to cheaper models, offloading or summarizing large tool results, capping tool call loops to prevent infinite loops, and trimming conversation history with a sliding window. These techniques help developers avoid wasteful token usage while maintaining agent performance.

[token-costs] [agents] [aws] [llm] [prompt-caching] [cost-optimization]

→ Watch on YouTube · → Full summary

[03] hermes 1 signal

MATTHEW BERMAN · 1D AGO

"The best thing since OpenClaw" (Hermes Tutorial)

Hermes is the successor to OpenClaw, offering a vastly simpler setup and a richer out-of-the-box experience with pre-installed skills, plugins, scheduled tasks, memory, and multi-provider support. The video demonstrates a 2-minute installation via Hostinger hosting, then walks through configuring skills, setting up Telegram integration, creating a daily brief automation, and showing self-healing behavior. The key differentiators are ease of use for non-technical users, a built-in skills marketplace, and the ability to bring your own inference provider.

[hermes] [openclaw] [ai-agents] [hostinger] [skills] [telegram]

→ Watch on YouTube · → Full summary

[04] local-models 1 signal

BIJAN BOWEN · 1D AGO

Ornith 1.0 First Look & Test – The BEST New Local Coding Models?

Ornith 1.0, a family of fine-tuned local coding models (9B dense and 35B MoE variants tested), shows promising performance on coding and 3D generation tasks, with the 35B MoE notably outperforming the 9B in most tests. The models leverage a self-improving training framework using GRPO, and the video highlights their potential as open-weight alternatives amid restricted access to frontier models.

[local-models] [coding] [fine-tuning] [ornith] [qwen] [open-weights]

→ Watch on YouTube · → Full summary

[05] gpt-5.6 1 signal

CALEB WRITES CODE · 1D AGO

GPT 5.6 explained in 8min..

OpenAI's fictional GPT 5.6 model is announced in three tiers (Saul, Terra, Luna) with pricing ranging from $1 to $5 per million input tokensasi. Chinese labs like Deepseek and Qwen are undercutting prices and catching up in capabilities, while US government regulation pressures frontier labs. The video also highlights a chip-level advantage with Cerebras enabling 750 tokens per second for the Saul model.

[gpt-5.6] [openai] [anthropic] [regulation] [chinese-models] [cerebras] [pricing]

→ Watch on YouTube · → Full summary

Frontier Notes · Generated Jun 28, 2026 · 6 of 6 signals
You received this as a Frontier Notes recipient.