Frontier Notes — Jun 25, 2026

Today

Google released a 51-page masterclass on agentic engineering, arguing the model is only 10% of the system while the harness (rules, tools, workflows, tests, guardrails) makes up the other 90%, a framework that reframes AI development around repeatable infrastructure.

Editor's Notes

This week's videos collectively underscore a paradigm shift in AI agents: from model-centric to infrastructure- and process-centric architectures. Several talks emphasized that reliability, scalability, and business value come from treating the agent's log as its identity (Omnara), using recursive verification loops (OpenProse), and investing in world models for synthetic training (Qwen-AgentWorld). Practical tools like Claude Code and Hermes Agent demonstrate how these principles translate into revenue generation and 24/7 automation.

Key Takeaways

Adopt the 'log is the agent' architecture: treat the append-only event log as the source of truth for agent state, enabling trivial forking, migration, and failure recovery.
Invest 90% of effort in building a repeatable harness (rules, tools, workflows) rather than chasing model improvements—Google's masterclass shows this dramatically reduces long-term costs.
Use recursive verification loops (e.g., Claude Code's 'roast' skill, OpenProse REPL) to self-critique outputs before delivery, turning agents from 'mismanaged geniuses' into reliable executors.
Leverage world models like Qwen-AgentWorld for synthetic RL training—hallucinated environments can outperform real-world training for self-reflection and adversarial robustness.
Upgrade agent interfaces with persistent memory and cron jobs (Hermes Agent) to enable 24/7 autonomous workflows like expense tracking, research, and analytics without manual prompting.

[01] llm 2 signals

AI ENGINEER · 14H AGO

Recursive Coding Agents - Raymond Weitekamp, OpenProse

Recursive coding agents, inspired by recursive language models (RLMs), unify tool calling and reasoning to make AI agents reliable outcome-deliverers rather than mismanaged geniuses. RLMs externalize the prompt as a variable and use a read-evaluate-print loop (REPL) to symbolically decompose problems, enabling small models like Qwen 3.59B to beat frontier models on long reasoning tasks. Applying RLM principles to coding agents—through dynamic workflows (Claude Code) or the OpenProse language—allows agents to recursively call themselves, verify sub-agent work, and capture golden sessions for repeatable reliability.

[llm] [agents] [recursive-language-models] [coding-agents] [openprose] [claude-code]

→ Watch on YouTube · → Full summary

SAM WITTEVEEN · 22H AGO

Qwen-AgentWorld The World Model for RL Environments

Qwen-AgentWorld introduces a world model that hallucinates environments to train agents more effectively than real-world RL training. The model predicts auto-regressive text outputs across seven domains (terminal, software engineering, web search, MCP tools, web browsers, desktop OS, Android OS) after receiving a state and action. This technique enables faster synthetic data generation and improved agent robustness through adversarial training and self-reflection.

[llm] [world-model] [reinforcement-learning] [quen] [agents] [synthetic-data]

→ Watch on YouTube · → Full summary

[02] openai 1 signal

LATENT SPACE · 13H AGO

Cooking with OpenAI’s Research Chief: AGI, o1, Evals, and Scaling Laws — Mark Chen

OpenAI's Chief Research Officer Mark Chen discusses the enduring power of scaling laws, the importance of replication for developing research taste, and the company's principled focus on pre-training, RL, and alignment as a stable high-level roadmap. He addresses the evals crisis, the need for new benchmarks, and the shift toward vibe research where models handle more execution. The conversation also covers leadership via meritocracy, the jagged frontier of model capabilities, and the underrated value of connecting research primitives to real-world agentic use cases.

[openai] [scaling-laws] [research-taste] [evals] [reinforcement-learning] [vibe-coder] [agi]

→ Watch on YouTube · → Full summary

[03] agents 1 signal

AI ENGINEER · 14H AGO

The Log Is The Agent - Ishaan Sehgal, Omnara

An agent's identity is fundamentally its log—the append-only event history capturing every input, output, and tool call—not the model or runtime, analogous to a video game save file. Treating the log as the agent simplifies reliability, scalability, forking, and migration, enabling workers to be disposable and sessions to survive failures. Current agent infrastructure treats logs as side effects, leading to fragility and lock-in, whereas a log-centric architecture makes these properties structural.

[agents] [logs] [architecture] [reliability] [scalability] [ownership]

→ Watch on YouTube · → Full summary

[04] claude-code 1 signal

NATE HERK · 15H AGO

I asked Claude Code to make me as much money as possible

Claude Code can be upgraded with four specific strategies to transform it from a productivity tool that agrees too readily into a revenue-generating business partner. The upgrades include a 'roast' skill that stress-tests ideas, a verification loop that checks work before delivery, context-management techniques to prevent performance degradation, and sub-agent workflows using /goal to parallelize tasks. These methods help users make better decisions, ship working products faster, and scale output without being the bottleneck.

[claude-code] [ai-agents] [prompt-engineering] [developer-tools] [automation] [business]

→ Watch on YouTube · → Full summary

[05] ai-assistant 1 signal

JASON LEE · 18H AGO

4 Insane Ways I’ve Been Using Hermes Agent

Hermes Agent is a 24/7 AI personal assistant with persistent memory, able to learn user preferences and execute tasks like expense tracking, research, and analytics. The video highlights four use cases: automatic business expense dashboard & Xero integration, daily SaaS idea generation from forums and trends, AI news curation, and YouTube channel analytics with competitor analysis. Setup is simplified via Hostinger's managed VPS, with Telegram/WhatsApp connectivity and cron jobs for scheduled tasks.

[ai-assistant] [agents] [productivity] [automation] [business-tools]

→ Watch on YouTube · → Full summary

[06] tv-deal 1 signal

DIGITAL SPACEPORT · 1D AGO

Stop! This 98" PRIME STEAL is $1,900

The TCL 98-inch QM7K TV is on a Prime Day deal for about $1,900, which is $300 less than what the host paid. It supports 144 Hz variable refresh rate for gaming, doesn't require an internet connection or account sign-in, and delivers near-perfect picture quality with only two settings adjustments. The main flaw is significant glare and reflection, especially in bright rooms.

[tv-deal] [gaming] [home-theater] [tcl] [prime-day] [large-tv]

→ Watch on YouTube · → Full summary

[07] ai-coding 1 signal

COLE MEDIN · 1D AGO

Google Just Dropped a Masterclass on Agentic Engineering (It's SO Good)

Google released a 51-page master class on AI-driven software development, framing AI coding as a spectrum from vibe coding to full agentic engineering, with the key insight that the model is only 10% of the system while the harness—comprising rules, tools, workflows, tests, and guardrails—makes up the other 90%. The article emphasizes that investing upfront in a repeatable harness (the 'factory' model) dramatically reduces long-term costs and inefficiencies, and introduces concepts like static vs. dynamic context, conductor vs. orchestrator modes, and token economics to guide practitioners.

[ai-coding] [agentic-engineering] [sdlc] [harness] [google] [developer-tools]

→ Watch on YouTube · → Full summary

Frontier Notes · Generated Jun 25, 2026 · 8 of 8 signals
You received this as a Frontier Notes recipient.