Google releases ADK Go 2.0 with graph-based workflows and human oversight

Google has released version 2.0 of its Agent Development Kit (ADK) for Go, introducing a significant shift toward graph-based workflow design for building reliable multi-agent applications. The update addresses the complexity of production agent systems that go beyond simple prompts to handle classification, branching, failure recovery, and human oversight.

What's new in ADK 2.0#

The core innovation is a graph-first architecture where applications are expressed as nodes connected by edges rather than ad-hoc control flow. Each node represents a unit of work, and the runtime scheduler handles concurrent execution, state persistence, and the ability to pause and resume workflows — even after process restarts.

The toolkit provides several built-in node types:

• Function nodes wrap typed Go functions with automatic input/output schema inference
• Emitting function nodes allow streaming events and human pauses without custom logic
• Agent nodes embed existing agents directly into graphs
• Tool nodes convert tools into workflow steps
• Join nodes gather outputs from parallel branches
• Dynamic nodes let developers write runtime orchestration in standard Go
• Workflow nodes enable nesting of entire sub-workflows as single steps
• Parallel workers run tasks concurrently across lists
• State-bound nodes pull session values directly into typed parameters

Edges between nodes can carry routing conditions, enabling sequential chains, conditional routers, fan-out and fan-in patterns, and even loops as first-class constructs. A common pattern uses an LLM agent as a "brain" to classify messages, with edges automatically routing to appropriate handlers.

Built-in human-in-the-loop#

Production workflows often need human approval or input at specific checkpoints. ADK 2.0 makes this a core feature — any node can pause execution and request a human response, which the workflow durably persists in session state.

Paused workflows can resume in two ways:

• Handoff: the human's response flows directly to the next node
• Re-entry: the paused node re-runs with the response available via context