March 25, 2026

TL;DR. Bet on what doesn't change — humans, UX, integrations, and agent feedback loops — not on benchmark gains the next model release will erase.

Takeaways

• "ngmi": context-trimming tricks, hybrid retrieval, finetuned cheaper-than-Opus models, custom memory systems, context graphs — all get steamrolled by the next frontier model.
• "wagmi": product/UI, distribution, integrations, fast CI and verification loops, background agent infra, and meeting users where they already work.
• Durable moats live in human-facing surfaces and infrastructure; transient moats live in model internals.

TL;DR. CLIs built for humans break agents — make every input a flag, every error actionable, and every command idempotent.

Takeaways

• Kill interactive prompts: if an agent can't pass it as a flag, it's stuck. Keep prompts as a fallback, not the default path.
• --help should lead with examples, not prose — agents pattern-match off mycli deploy --env staging --tag v1.2.3 faster than they read descriptions.
• Idempotency + --dry-run + --yes are non-negotiable: agents retry constantly, need to preview destructive actions, and should bypass "are you sure?" cleanly.
• Return structured data on success (deploy ID, URL, duration) — skip the emojis, agents parse the next step from your output.