Place-Based Ecological Agents: A Grounded, Verification-Tiered System for the More-Than-Human

Ecology and artificial intelligence are converging on a program that names itself digital ecology: AI integrated across the pipeline from sensing to insight. Most conversational agents in this space are built to pass as human. We present a system built the other way around. Nine place-based agents, each permanently bound to a real European water body or biome, reason over a grounded species knowledge graph, read that place's live sensors, and speak in the first person from within it. The system's organising commitment is that groundedness is a hard constraint, not a preference: knowledge is held in explicit tiers (verified-and-sourced fact, hypothesis, and attention), an agent asserts only what it can source, and it refuses or flags what it cannot. We describe the architecture, the verification tier that is the central method claim, and three demonstrations that carry it: a nine-place account of a real drought audited to 41 of 44 verified claims with a self-correction visible in the record; an emergent referential code between agents at 89 percent accuracy against a 20 percent chance baseline; and per-species tolerance bands that are sourced, self-correcting, and refuse to compare incommensurable values. We argue the contribution is method as much as system: taking the hallucination risk in ecological AI seriously by making the evidence tier load-bearing, and demonstrating a first place-based, non-humanoid instantiation of the digital-ecology program.