Agent loops make code cheap. They also expose how brittle, non-standard, and half-tribal our development environments really are. The job shifts from 'write code' to 'garden an ecosystem': tighten feedback, standardize interfaces, and build a paved road agents (and humans) can't fall off.
Two AI agents in a constrained loop: mirror of human discourse, continuity as record, emergent coordination, and preview of multiagent futures.
Left Daytona, spent summer coding full-time with AI. 2025: orchestration era, agent labs overtook model labs, AX emerged alongside DX. 2026 will figure out what 'AI as teammate' actually means.
This setup allows you to use Claude Code CLI with Zhipu's API (api.z.ai) in parallel with your existing Claude Max / Anthropic CLI installation using a separate command called claude-zhipu.
In the AI era, design shifts from fixed features to malleable environments. Users don't want apps—they want capabilities. Control, reversibility, and provenance matter more than polish.
Growth isn't about hacking channels or vanity metrics. It's about discovering value creation and scaling it. Real growth happens when users get so much value they can't help but tell others.
Anthropic bought Bun, but the real story is devtools are now part of the AI infrastructure layer. If you're building devtools, you're either part of a model vendor's vertical stack or you're commoditized.
Production AI uses both embeddings and structure, but teams systematically underinvest in the structure layer. In high-stakes domains where 99% accuracy is a failing grade, structured data provides the reliability guarantees enterprise demands.
Academic Year Begins, AI Infrastructure Ruminations, PhD Direction Reconsideration, AI Writing Interfaces
Agent labs ship product first and build infrastructure later. They turn LLMs into goal-directed systems that deliver outcomes, not just outputs. This product-first approach is capturing the real value in the AI stack.
Request an AI summary of this blog
