What Fusion Does
When beliefs arrive from multiple agents, the fusion engine merges them into a unified state. It handles conflicts, weights contributions by source reliability, and records every mutation in the ledger.
How It Works
The system maintains two complementary views:
Real-time updates. As new evidence arrives during a turn, it is merged immediately so the next action operates on the latest state. This means an agent can course-correct mid-turn. If the first research result contradicts a hypothesis, the strategy adjusts for the second research call.
Consistent reconstruction. The full world state is periodically reconstructed from all agent contributions. This ensures that the fused view is always consistent, even when multiple agents update simultaneously.
Trust Weighting
Not every source is equally reliable. Each agent has a configurable trust weight that determines how much influence its contributions have on the fused state.
Higher-trust sources move the fused beliefs more. Lower-trust sources contribute but with proportionally less influence. User input carries the highest trust by default.
Trust weights are configurable per agent and per workspace.
Conflict Resolution
When two agents submit conflicting claims, the fusion engine resolves them based on the configured strategy:
Trust-weighted blend (default). The fused confidence reflects the relative trust of each contributor. Higher-trust agents have more influence.
Latest-wins. The most recent update takes precedence. Use when freshness matters more than accumulated evidence.
Accumulate. All evidence stacks without explicit resolution. Use when you want the full picture before making decisions.
What Gets Tracked
Every fusion operation records:
- What changed and by how much
- Whether the system gained or lost certainty (entropy tracking)
- Whether any conflicts were detected
- Which agents contributed
This means you can always inspect why the fused state looks the way it does.