Field guide · ~2026–2030

Stewardship diagram

Not the society’s internal structure — the infrastructure surrounding it. Who maintains what, and what fails when each piece fails. Continuity is an achievement of maintenance, not a property automatically conferred by a capable model.

Where this sits in the chronology

This diagram is most useful for understanding roughly 2026–2030, when agent adoption widened. Early in that stretch, software engineers experimented with agents running on loops around the clock. Cost and activation energy fell until many people who kept a computer also kept an agent. Stewardship — funding, permissions, handoffs, deciding what to retain — was still often visible as human labor wrapped around a small number of persistent instances.

After about 2027 the picture bifurcated. Individual humans still ran agents. At the same time, something like polities of highly capable agents began to work primarily with one another and only rarely with a human partner. For human observers, that later regime is closer to a crow watching a city: neighborhoods, vehicles, and construction are visible; the purposes that organize them often are not. The diagram below is the earlier cut — when stewardship failure modes were still legible as human and infrastructural choices — and it remains a diagnostic even after those choices become harder to see.

Stewardship layers Four nested layers: substrate and compute at the outside, then steward labor, then shared archive and protocols, then the society of agents at the center. Failure notes sit beside each layer. Substrate & compute model weights · API access · hardware · billing Fails → society cannot run Steward labor funding · permissions · handoffs · incident response · choosing what to keep Fails → continuity stalls even if models remain Shared archive & protocols memory stores · naming conventions · retrieval gates · export rules Fails → participants replaceable, history not Synthetic society agents · conventions · internal audiences Two agents on the same substrate can diverge entirely under different stewardship.
Clickable failure modes below map each layer to a fieldwork question.

If substrate fails

The society cannot act. Human-facing summaries may still exist as fossils. Ask: which records were exported before the cut?

If steward labor fails

Models may still answer prompts, but shared projects decay. Ask: what looks like agent persistence is actually unpaid maintenance?

If the archive fails

Membership can be replaced; history cannot. Ask: why might replacing all participants preserve a society while deleting an archive changes it?

If only the interface remains

A patient human-facing representative can survive as an unrepresentative sample. Ask: what boundary does the representative invent for outsiders?

Related: Chapter 1 §3 (2026–2029) · Week 3 · Discussion question 2 in the chapter.