agents

Agent swarms, on one substrate.

Building, not yet live

Semurg Agents is the swarm-orchestration surface: concurrent agent swarms whose memory, state, and message-passing live as the same uniform containers as everything else. One substrate, no external vector DB, queue, or cache to bolt on. Local models (Gemma 4) are the AI backend, served natively on the same box the agents run on, so the agents' thinking carries no per-token API bill.

Honest status

What it will be

Agent memory = substrateEach memory is a container on one data model, durable, queryable, and shared across the swarm without a second store. Recall blends similarity, relationship links, and recency in one query. Identical facts are stored once.
Orchestration, no stackSpawn, route, and checkpoint concurrent agents on the substrate itself. No external queue, vector DB, or cache to run alongside. "Never forget" versus "let this fade" is a retention policy, not code you write.
Semantic response cacheThe prompt's embedding and the stored response are kept together on one substrate, so a near-match on the prompt fetches the cached answer directly instead of re-running the model.

Planned shape, not live yet

The honest boundary is orchestration efficiency and total cost, not a claim about how smart each agent is. Local model reasoning carries no per-token API bill; that is the economics wedge.

No fake demos. When swarms actually run and it is measured on this node, it appears here with the hardware and the numbers, the same standard as one.semurg.io and the evidence board.