January 24, 2026

TL;DR. You get ~12 career "shots" of ~4 years each, and timing (天) beats place (地) beats people (人) for whether they land.

Takeaways

• Reframe your career as 12 swings, not a continuous grind — each 4-year window is a discrete bet.
• Timing > environment > people, in that order. Riding the right wave matters more than who you know.
• Implication: when a shot misses, the lesson is usually about timing, not effort or talent.

TL;DR. Two years building AI agents for finance: sandbox everything, ditch RAG for agentic filesystem search, and assume the model will eat your scaffolding.

Takeaways

• "The fear of being wrong becomes our best feature" — financial agents demand paranoid precision because one bad DCF number destroys trust forever.
• Fintool retired its embedding pipeline and went all-in on Claude Code's filesystem-first approach (ReadFile/WriteFile/Bash + S3); the author's "RAG obituary" post called this shift before it was consensus.
• Design for obsolescence: build scaffolding you expect newer models to render unnecessary, and lean on Temporal for long-running tasks plus domain-specific evals to catch errors before they cost money.