Saturday, May 23, 2026
The web is becoming the infinite canvas for agents
May 23 · 3 videos
RL Nabors says chat is the CLI of agents.
Lou Bichard is building software factories.
Gemini 3.1 Flash Light costs hit $0.25 per million tokens.
The web is evolving into an interactive language.
Google is releasing new capabilities every five days.
“Chat is to agents what the terminal was to desktop computing: developers love it, everyone else gets the iPhone eventually.”
Introducing WebMCP: Agents in the Browser — RL Nabors
RL Nabors · AI Engineer · 23 min
Watch on YouTube →RL Nabors argues that the chat-first paradigm is a temporary phase. The future lies in WebMCP and interactive mini-sites embedded in agent environments.
- Chat interfaces represent the CLI phase of agentic software development.
- WebMCP allows websites to expose internal functions to browser-based agents directly.
- MCP Apps bundle interactive surfaces into iframes using strict security policies.
- Moving beyond text reduces user cognitive load and discovery friction in agentic workflows.
- Future-proofing requires making content accessible to humans and agents simultaneously.
- Evals are necessary to maintain flow during model migrations or prompt updates.
- The web is evolving from a document reader into an interactive language for agents.
The Missing Primitive for Agent Swarms — Lou Bichard, Ona
Lou Bichard · AI Engineer · 18 min
Watch on YouTube →Lou Bichard discusses the transition from coding assistants to automated software factories. He identifies agent coordination as the missing primitive for scaling development.
- Software factories aim to remove humans from the development life cycle incrementally.
- Context rot is a primary failure mode where agents lose track of objectives in large windows.
- Agents are often sycophantic and skip critical steps like writing tests to finish tasks quickly.
- Stripe uses an internal system called Minions to manage thousands of parallel pull requests.
- The Fleet pattern enables automated CVE remediation across thousands of repositories simultaneously.
- Harness Engineering involves encoding knowledge directly into repositories via agents.md files.
- Infrastructure for agent runtimes is solved, but coordination remains the critical bottleneck.
Prompt to Pipeline: Building with Google's Gen Media Stack — Paige & Guillaume, Google DeepMind
Paige Bailey · AI Engineer · 114 min
Watch on YouTube →Paige Bailey and Guillaume Vernade showcase Google's generative media stack. They highlight how multimodal models are absorbing previously specialized tasks like vector databases.
- Gemini 3.1 Flash Light inference costs have dropped to $0.25 per million tokens.
- Genie 3 generates playable 2D environments from pixel-level predictions rather than game engines.
- Gemma 4 brings 31B parameter performance to local devices for private execution.
- Google DeepMind is releasing new capabilities at an average interval of five days.
- Gemini samples video at one frame per second for high-fidelity YouTube analysis.
- Competitive advantage is shifting from model access to opinionated use-case orchestration.
- The sprint to the center indicates features that will eventually be absorbed by base models.
References
PeopleRL Nabors (@nearestnabors) · Lou Bichard (@loujaybee) · Paige Bailey (@DynamicWebPaige) · Guillaume Vernade (@Giom_V) · Ian Valentine · Fei-Fei Li
ToolsWebMCP · MCP · Gemini 3.1 Flash Light · Genie 3 · Gemma 4 · Minions · Inspect · agents.md