State Management

The State Problem

When an agent crashes, its memory (RAM) is gone. If you were storing conversation history in a Python list, it's lost forever.

Persistent Context

Consonant maintains the Workflow State in the Control Plane (backed by Etcd/Postgres). When the Control Plane calls an Agent, it passes the relevant Context (conversation history, previous tool outputs) as input.

// Input passed to Agent
{
  "goal": "Summarize this",
  "context": {
    "file_content": "...",
    "previous_summary": "..."
  },
  "state_id": "run_123"
}

The Agent performs its task and returns the State Update. It does *not* persist state locally.

// Output from Agent
{
  "result": "Here is the summary...",
  "state_update": {
    "summary_completed": true
  }
}