Skip to content

SwarmPlane

A living task graph for agent orchestration on Google Cloud.

Agent frameworks fake one of two halves.

Plan-and-execute designs assert the whole future up front, then flail when step three returns something they didn't anticipate. ReAct loops adapt beautifully but leave behind a transcript — not something you can inspect, diff, or hand to an auditor.

SwarmPlane grows the graph from outcomes. Expansion is append-only: a retry is a new node whose parent is the failure, never an edge back to a completed one. The graph stays acyclic, and the complete causal history is preserved by construction.

graph TD
    T[Task: monthly revenue by region] --> S[Inspect schema]
    S --> Q1[Draft SQL v1]
    Q1 --> F[Failed: no column 'region']
    F --> D[Discover: region lives in dim_store]
    D --> Q2[Draft SQL v2 with join]
    Q2 --> R[Validated result]

    classDef fail fill:#7f1d1d,stroke:#ef4444,color:#fff
    classDef ok fill:#14532d,stroke:#22c55e,color:#fff
    class F fail
    class R ok

The failure is not erased and retried in place. It stays in the graph as the parent of the discovery it caused. That is the whole idea.

Two properties follow

The execution graph is the audit log. Every node is a span. Every edge is a causal claim. There is no separate trace to reconcile against state — the thing that ran and the record of what ran are one object.

Every seam is an open protocol. MCP for tools, A2A for agents, A2UI for interface. Nothing proprietary sits between layers — including SwarmPlane itself, which you can remove and keep your agents.

Three opinions

  1. The graph is primary, not the agent. Agents are node executors.
  2. Every boundary is an open protocol. No proprietary abstraction at any seam.
  3. GCP-native, not cloud-agnostic. Portability is not a goal.

SwarmPlane is not unopinionated. It is transparent: every decision is visible, inspectable, and replaceable.

On the name

"Swarm" here means fleet, not emergence. A plane coordinates workers against a shared graph. Nothing about this design is non-deterministic by intent.

Why append-only makes serverless work

Cloud Run caps a request at 60 minutes. Frameworks that model an agent run as one long process fight that ceiling. SwarmPlane has no long process: each graph expansion is one short request, the graph is durable, and a crash resumes from the frontier. The constraint that breaks other designs is the one this design is shaped for.

Status

Pre-alpha. SwarmPlane is a reference implementation — a small, deliberately readable codebase meant to be read and copied, not imported. There is no API-stability promise.

Start with The Living Graph, which is the argument everything else rests on, then Prior Art for an honest account of what is and isn't new here.