Topology and Brews

Forge can run a simple single-model agent, but the runtime also supports multi-model topologies and graph-shaped execution plans.

These surfaces are Rust-first Wave 2 capabilities. They are documented here as the reference contract; cross-language SDK parity follows the conformance program.

Model topology

A ModelTopology is the set of model slots available to an agent or Brew. The single-model case is represented as a topology with one default slot.

Each slot can carry:

• role name, such as default, planner, coder, or critic
• primary provider reference, such as anthropic:claude-sonnet-4-5-20250929
• ordered fallbacks
• required runtime capabilities
• latency and cost preferences
• optional Arsenal scope narrowing

The router chooses a slot; the topology only declares what is available.

Provider references

Forge uses the namespace:model provider reference pattern across the runtime. That lets a topology move between providers without changing agent logic:

anthropic:claude-sonnet-4-5-20250929
openai:gpt-4o
google:gemini-2.5-pro
local:llama

Provider capability negotiation decides whether a slot can satisfy a request. See Provider Capability Matrix.

Brew graphs

A Brew is a directed graph of typed nodes and edges. It generalizes flat sequential, parallel, and router workflows into a single graph model.

Common node kinds include:

Before durable execution, a dynamic Brew is frozen into a ResolvedBrewPlan. Flowers executes resolved plans, not raw symbolic Brews.

Status

Use this surface when one agent needs different model roles, explicit fallback behavior, or a workflow graph that can be inspected and replayed.