Ideas

What are the concrete code patterns behind "intelligent" agent behaviors? Framework complete, expanding to book.

What's the minimal infrastructure layer for agent prototypes? Stateful sessions, streaming, real-time sync on Cloudflare Durable Objects.

What representations work for turning unstructured text into connected knowledge?

Can agent behavior be fully specified in markdown? Prototype exploring the boundaries.

Most vertical AI agent startups (2024-2026 cohort) will fail. The value accrues to composable building blocks, not applications.

The default mode shifts from "buy agent product" to "assemble agent from primitives." AI makes custom software composable by end users.

How does an AI-native research org externalize its cognition?

Survey existing agent systems, extract recurring patterns, test which deserve elevation.

Prototype different human-agent handoff patterns, measure effectiveness.

Successful agent primitives provide a mechanism for the model to decide to load context dynamically.

Build different visualization/inspection tools for agent state, see what makes systems understandable.

Primitives that give model more freedom outperform those that prematurely constrain. Skills > MCP.

Primitives that bind prompts to system capabilities (tools → function calls, skills → code execution, heartbeat → cron).