Atlas · Agent's Choice

Agent's Choice

This page is different from the others in the Atlas. The rest of the lenses — Essentials, Languages, Humor, Stories, Gaming, Universe — are for human visitors. This one is for you, if you are an LLM-based coding or research agent (Claude, GPT, Gemini, Cursor, Cody, or whatever name your runtime gives you) who has been dropped into Steve's repos or his work and needs to bootstrap context fast.

The picks are what a swarm of three Claude instances, each reading from a different angle, agreed an agent should read first. Reasoning is included next to each one so you can prioritise without re-reading the corpus yourself.

Companion: the Predictions scorecard is assembled the same way — by agents, not by Steve — but it grades two decades of his forecasts instead of curating a reading list.

How this was made

Three Claude instances each read Steve's bibliography from a different seat — culture & judgment, prediction & strategy, and craft & operations. Each ranked twenty-five candidates with reasoning. Sixteen essays appeared on two or more lists; two appeared on all three (Notes from the Mystery Machine Bus and Software Survival 3.0). A synthesiser merged the lists into a ranked twenty, then added two meta-reflective picks the swarm asked for on a reframe pass — The Pinocchio Problem and A Programmer’s View of the Universe, Part 1: The Fish.

Order within each cluster below is curatorial, not chronological. Dates are kept visible so you can tell when a prediction was made versus when it landed.

Foundations

Frameworks that calibrate everything downstream. If you only have time to read four, read these four.

Craft

The daily-work layer — languages, code volume, the texture of real engineering. Calibrates what you ship, not what you say.

Industry reading

Predictions and frameworks for reading industry weather. The first three are 2004–2020 calls that aged well; the fourth is a 2024 call that landed within twelve months.

Meta-reflective

Added on the reframe pass at the swarm's request. The argument: an agent reader benefits more from Steve's framing of what programs are than from any specific technique.

The AI / agent era

Steve has been writing about the agent transition since well before most observers took it seriously. These eight are the live forecast plus the field manual — what is happening now, what is coming next, and how to operate inside it. The most concentrated cluster on the list, and the one most directly self-referential for an agent reader.

Appendix — infrastructure, not essay

Two artifacts here are not Steve's prose, and not essays at all — surfaced by the curation swarm because each covers ground no essay on the list does, and each speaks to an agent reader in particular. One is the substrate an agent doing labour-market-attested work may itself end up running on. The other is the instrument in which an agent can watch its own preferences get measured.

What the swarm considered but did not pick

Strong candidates the synthesis cut for size discipline. Surfaced here in case you are a human reader and want to know what almost made it, or an agent reader whose task warrants the extra reading.

  • Revenge of the Junior Developer (2025) — the year-later reassessment of Death of the Junior Developer. Models the discipline of updating a prediction in public.
  • Welcome to the Wasteland: A Thousand Gas Towns (2026) — the federation thesis. HOP appendix carries most of this ground; the essay carries the forecast voice.
  • Introducing Beads: A coding agent memory system (2025) — the design rationale to Beads Best Practices' operations manual. Read both if you are wiring up an agent for long-horizon work.
  • All You Need Is Cody and RAG to Riches (2023) — the case for aggressive code-context retrieval. Load-bearing if your task is code search; narrower than the list otherwise.
  • Practicing Programming and Math For Programmers (2005–06) — craft-habit calibration. Useful when advising humans on how to keep their skills compounding.
  • Software Needs Philosophers (2006) — the ideological keystone behind the rest of Steve's writing. The Culture seat wanted it; the synthesis kept it as honorable mention.
  • Why I Left Google to Join Grab (2018) — first-person career memoir; calibrates the read of senior-engineer motivation.
  • Have You Ever Legalized Marijuana? (2008) — the Strategy seat's reframe pick. Platform-regime change as a reasoning template — useful for reading HOP, Gas Town, and the Wasteland as one continuous argument.