The Architecture of Modern AI Agents

Philipp Schmid delivered a masterclass in agent architecture today, breaking down several key concepts that are reshaping how we think about AI systems.

"Subagents are specialized AI agents. They are most of the time used in combination with an orchestrator, which delegates tasks to them. A subagent is just like a normal agent and has the same components."

This delegation pattern mirrors how effective human teams operate—a coordinator who understands the big picture, with specialists who go deep on specific problems. The shift from "Shallow Loops to Deep Agents" (as Schmid frames Agents 2.0) represents a fundamental change in how we architect AI systems.

"The discipline of designing and building dynamic systems that provides the right information and tools, in the right format, at the right time, to give a LLM everything it needs to accomplish a task."

This framing moves us beyond prompt engineering into systems thinking—it's not just about what you say to the model, but about the entire information architecture surrounding it.

From Theory to Practice: Building Agents

Philipp Schmid also released a practical guide for building AI agents with Gemini 3 Pro:

"Construct a working prototype in under 100 lines of code. Start from basic text generation to a functioning CLI Agent."

Meanwhile, Rohit outlined five agentic AI projects that actually get developers hired:

1. Agentic Workflow Automation (auto-generates tools on the fly)

2. Memory-Driven Customer Intelligence Agent

3. Self-Correcting Multi-Agent Researcher

The pattern is clear: the market wants builders who can create agents that learn, adapt, and correct themselves.

Local LLMs Hit Browser-Native Milestone

Surya Dantuluri shipped something remarkable—Qwen3 0.6B running entirely in the browser via WebGPU:

"No installation or servers necessary and runs offline. Available forever for free and open source."

This represents a significant democratization moment. When capable models run in a browser tab with no backend, the barriers to AI experimentation effectively vanish.

The Economics of AI-Powered Micro-Companies

Greg Isenberg made a bold prediction:

"We're about to see the largest boom in micro-companies in history. 1 person to 10-people businesses that generate real cash, serve tiny but passionate communities, and operate with leverage (organic social, AI etc) that used to require mega teams."

This thesis is being validated in real-time by developers like 0xSero, who's running minimax-reap-162B locally on 8x 3090s:

"3500 tps for prompt processing, 50~ tps for prompt generating. Running in Claude Code, used MCPs hooks, subagents, skills, all perfectly faster than Claude."

When individuals can run frontier-competitive models on consumer hardware, the traditional moat of compute access erodes quickly.

Rethinking Developer Workflows

Geoffrey Litt proposed a workflow inversion that's worth considering:

"Instead of having Claude Code make a PR, ask it to output a Markdown tutorial doc + build-it-yourself."

This "tutorial doc" approach transforms the AI from a black-box code generator into an educational partner. You understand what's being built because you're building it with guidance.

Santiago on code reviews:

"AI can check your code 100x faster and catch 10x more issues than anybody can. There are two remaining reasons for a manual code review: 1. To transfer knowledge within a team 2. To ensure AI didn't miss critical aspects."

The role of human review is shifting from "find bugs" to "transfer context and verify judgment."

Open Source Finance Gets Serious

Virat announced Dexter, an open-source deep research agent for finance:

"The future of finance isn't closed. It's open source. It's crushing evals, improving fast, and every line of code is yours."

Financial AI has traditionally been locked behind proprietary walls. Open alternatives could reshape who gets access to sophisticated financial analysis.

Tools of the Day

• Nano Banana Pro continues to impress for visual generation—multiple users showcased infographics and visualizations generated from single prompts
• OpenAI shipped ChatGPT Apps SDK UI components, making it easier to build custom ChatGPT integrations
• Computer control for AI agents via new tooling shared by Tom Dörr
• OpenAI's AI-Native Engineering Team guide dropped, covering how coding agents fit across planning, design, and maintenance phases

The Bigger Picture

Today's discourse reveals an ecosystem rapidly maturing past the "can AI code?" question into "how should AI systems be architected?" The answers emerging—subagents, context engineering, deep rather than shallow loops—suggest we're entering a phase where the craft of AI system design becomes as important as the underlying models themselves.