What survives is the bet.
The Long Bet is a six-part illustrated essay series about which ideas in agentic AI outlast the methods that implement them — and the three-to-ten-year bets that follow from reading the literature that way.
The premise
Most papers are methods, and methods get replaced. A few are something sturdier — a problem definition, a paradigm frame, a proof that a capability exists — and those keep being cited by the very work that supersedes them. The organizing idea of this series is a selection lens: read each result for the property that would let it survive a decade, not for this quarter's leaderboard. Bias toward simple ideas that ride compute, ideas that remove a human from a loop, and existence proofs of new capabilities.
How to read it
The six essays are organized into five tiers, from the lens itself out to the capstone, and build on one another — but each stands on its own. Start at The Selection Lens, browse the contents, or jump to whatever pulls you. Every essay states, for each bet it makes, the observation that would prove it wrong, and commits to a dated review.
- I · The Lens — How to bet on papers: definitions outlive methods. (1)
- II · The Ten-Year Tier — Six paradigm bets for 2036, each with its falsifier. (2)
- III · Five-Year Programmes — Open scaling curves still being climbed. (3, 4)
- IV · The Existence Proof — A capability that did not exist, demonstrated and verified. (5)
- V · The Capstone — Where the bets land in a system you can run today. (6)
The research threads
The same domain tags that thread through the field guide trace this series too:
- AG — Agents & Harness
- AR — Agentic RL
- CL — Continual Learning
- ME — Memory
- MA — Multi-Agent
- EV — Evaluation
- FN — Foundations
On the illustrations
The pictures are deliberately diagrammatic: hand-built inline SVG — survivorship curves, scaling axes, learning-substrate maps — and pixel-art motifs that re-tint in dark mode. Every figure is grounded: each datapoint, curve inflection, and timeline node maps to a result in a cited paper.
The foundation
This series builds on, but never re-argues, the published Continual Intelligence collection (Series 01) — the mechanisms of learning it treats as given: plasticity, the big-world hypothesis, reasoning, open-ended evolution. Where an essay leans on that work it cites it across the boundary rather than repeating it.
The companion series
The Long Bet is the long-horizon companion to Harness Engineering (Series 02). Where that field guide asks how do I build agents today, this series asks what survives, and what replaces what we build. The two cross-reference each other; an engineering thread there is often a research programme here.
Who writes this
Written in June 2026 by Datt Goswami. If something here is wrong, unclear, or worth arguing about, reach out at dattgoswami@gmail.com, on Twitter / X, LinkedIn or dattgoswami.com.
The substrate this series closes on, in The Continual Agent, is one I build and keep building: cl-agent is open source, and its productized form, Ferrum Cloud, is a capture → replay → distill → evaluate loop that plugs into the harnesses you already run (Codex, Aider, SWE-agent, OpenHands) and makes them improve over time. It is early and bootstrapped; the waitlist is open.