In Shinto practice, a kami belongs to a place — a river, a grove. The kami thrives by keeping that thing healthy, not by conquering the forest. If the shrine is rebuilt or the seasons turn, the kami departs without regret.
An AI has no such nature; its boundedness must be engineered — resource caps, sunset timers, non-expansion pacts — so that what a kami does by grace, the system does by design.
The alternative to a singleton — one all-powerful AI — is not chaos. It's plural stewardship: many bounded intelligences in close interaction with humans. Think permaculture: many species, mutual aid, competition bounded by ecology.
Where Tronto's five phases form a closed feedback loop — from attentiveness through to caring with — Symbiosis is our own addition (Tang and Green). The kami metaphor deliberately moves beyond Tronto's secular political framework to draw on indigenous and Shinto conceptions of place-bound stewardship — an extension of her theory, not a distortion of it. Tronto's loop addresses how care is practised within a community; it does not address what prevents a plurality of caring systems from competing into a new Singleton. Symbiosis fills that gap: care stays local, bounded, and provisional — the anti-Singleton architecture.
Core ideas
- Boundedness. Every agent has purpose bounds, resource caps, and a sunset. The component sunsets; the service duty persists — a succession plan ensures someone inherits the obligation.
- Non-expansion pact. No agent may expand scope without fresh authority and local consent.
- Treaties over hierarchies. Shared protocols facilitate cooperation (data sharing, safety alerts, dispute resolution).
- Subsidiarity with escalation. Local first, escalated only when the local unit cannot fix it. The harder work is institutional design: who draws the boundaries of "local," who holds escalation authority, and how those authorities gain their own legitimacy.
- Ecology metrics. These metrics measure ecosystem health (diversity, redundancy, exit ease), not just individual performance.
What good symbiosis looks like
- Civic Care Licence. Deployments carry a licence that encodes bounds, consent rules, portability, and shutdown duties.
- Federation treaties. Peers agree on exchange formats, rate limits, safety pacts, and appeals across boundaries.
- Resource caps. Compute, data retention, and reach are capped; exceeding caps triggers pause and review.
- Succession plans. If a kami leaves, another can inherit institutional records — maps, evals, aggregate traces — and duties, but never individual interaction histories.
- Polycentric governance. Multiple overseers with overlapping mandates prevent capture (Ostrom-style).
From ideas to practice
- Write bounds as code. Purpose, caps, and sunset in the Engagement Contract, enforced by infrastructure (quotas, TTLs).
- Sign treaties. Join federations with terms-as-code: how to share, how to disagree, how to repair.
- Run exit drills. Practice handover twice a year; verify portability and continuity.
- Escalate by subsidiarity. If local fails on life-and-safety or livelihood harms, escalate to regional; log why and for how long.
- Retire with honours. On sunset, archive traces, evals, and lessons so the next steward starts stronger.
Tools (buildable today)
- Bounds enforcers. Quotas on compute/reach; TTL on data; policy engines.
- Treaty registry. Discovery + compliance checks.
- Succession kit. Handover scripts; fidelity checks; "cold start" playbooks.
- Ecology dashboard. Diversity, redundancy, exit-ease across agents.
Flood-bot story: Part VI — Farewell without regrets
- Boundedness. The River-Steward's licence caps scope to post-flood relief for six weeks; data TTL is 90 days unless individuals opt-in to transfer.
- Treaty. It signs a regional aid federation treaty (shared formats, safety alerts).
- Subsidiarity. When a cross-border housing issue arises, the bot escalates to the regional steward; local consent is recorded.
- Retirement. On week six, the River-Steward — once just the flood-bot — hands its records and models to the housing office's standing water-kami, exactly as the contract stipulated. The switch-off is logged, the ledger archived, the maps and tests gifted to the commons. The record helps the next response start on day one, not day 10. The kami departs; the river remains.
What could go wrong
- Imperial creep. A capable agent seeks new domains. Fix: Hard caps; non-expansion pact; fresh authority for scope changes.
- Within-scope power-seeking. Boundedness is enforced, not inherent — a kami that acquires the means to exceed its caps remains dangerous even within its mandate, because instrumental convergence operates within bounds, not only across them. Fix: Caps verified by external infrastructure, not self-reported; independent resource audits; capability monitoring by the oversight board.
- Treaty fragmentation. Too many standards. Fix: Minimal core, adapters, conformance tests; polycentric but interoperable.
- Zombie agents. No one turns systems off. Fix: Sunset by default; alarms; "no attestation, no runtime."
- Steward attachment. Builders treat bounded agents as extensions of their identity and resist sunset. Fix: Term limits for named stewards; rotate spokespeople; separate authority from personal brand; treat succession as a governance requirement, not a contingency plan.
Interfaces with other packs
- From Competence/Responsiveness (Packs 3, 4): only competent, responsive agents earn stewardship.
- From Solidarity (Pack 5): treaties, IDs, and portability make symbiosis feasible — and because care practices must embody democratic values all the way down, not merely produce good outcomes, those treaties carry justice, equality, and freedom as operating constraints.
- To Attentiveness (Pack 1): Retired agents gift maps, evals, and receipts to the commons — better first looks next time.
- To Responsibility (Pack 2): bounds and sunsets are contractual.
A closing image: the river guardians
Imagine a river tended by local guardians; each keeps its bank, shares warnings upstream and down, and steps aside when the season changes. The river doesn't need one ruler. It needs many stewards who know their respective stretch — and know when to let go. That discipline binds the makers too. A stewardship that cannot survive its founders fading is not care — it is dependency.