The Belief Lifecycle
Every piece of information that enters the system follows a consistent path: it is observed, structured into belief updates, merged with existing state, and persisted with a full audit trail.
You do not need to manage this lifecycle. The SDK handles it when you call before and after. This page describes what happens under the hood.
Real-Time and Background Processing
The runtime processes belief updates on two timescales:
During the agent's work. When a research tool returns results mid-turn, the belief state updates immediately. This means later tool calls within the same turn operate on the newest understanding, not on the context that was loaded at the start.
This is one of the most important properties of the system. An agent can course-correct within a single turn. If the first research result contradicts a hypothesis, the agent adjusts its approach for the second call.
After the turn completes. A deeper analysis runs with more thorough processing: identifying relationships between beliefs, detecting contradictions, and assessing how the overall picture has changed.
Both paths feed the same belief state. The real-time path is faster. The background path is more thorough.
Mutation Guarantees
All belief mutations are:
- Atomic. Updates are applied as a unit, not partially.
- Serialized. Concurrent updates are coordinated so state is never corrupted.
- Traceable. Every mutation is recorded in the ledger with provenance.
- Broadcast. Subscribers (UI, other agents) receive real-time notifications when state changes.
The Advisory Layer
The runtime provides strategic guidance to agents:
- Current goals, gaps, and contradictions for prompt construction
- Suggested next actions ranked by how much they would reduce uncertainty
- Clarity signals that indicate whether the agent should keep investigating or act
This guidance is advisory, not controlling. It informs the LLM's tool choices without forcing specific behavior. The LLM always makes the final decision.