HC11 · Human Control & Oversight

Conflict Resolution Surface

Purpose Presents contradictory outputs from two or more agents and requires a human decision before the run can continue or close. Surfaces the conflict explicitly, showing which agents produced which outputs and the downstream consequences of each resolution path.

Use when Two or more agents in a network return contradictory results — different answers to the same question, conflicting recommendations, or incompatible data — and the run cannot proceed without resolving which output to use. Required before any downstream agent acts on conflicting inputs.

Do not use For minor output variation that doesn't create a decision conflict. Do not auto-resolve conflicts by choosing the most recent output, highest-confidence output, or any other heuristic without explicit human review. Auto-resolution is never acceptable.

Failure mode prevented Silent conflict resolution — the system choosing between contradictory agent outputs without surfacing the conflict to the human, hiding the disagreement behind a single merged answer.

Component UI

Component UI coming soon

Overview

Design rule

When agents disagree, the disagreement belongs to the human. Auto-resolution is never acceptable. The conflict must be surfaced, attributed, and resolved by explicit human decision before the run closes.

The Conflict Resolution Surface is triggered when the orchestrator detects that two or more agents have produced outputs that cannot be reconciled without a human decision. The run pauses at this point. The component shows: each conflicting output attributed to its source agent, the nature of the conflict (factual contradiction, recommendation divergence, incompatible data), and the implications of choosing each resolution path.

This component is not a confidence comparison tool — it does not recommend which output to prefer. It is a disclosure and decision surface. The human must make the call, and that decision must be recorded in the audit trail as a human governance event, not an automated resolution.

Anatomy

Conflict header: Conflict type, which agents are involved, timestamp, run status (paused)
Conflicting outputs: Side-by-side or stacked view of each agent's output, attributed by name and agent ID
Conflict description: Human-readable summary of what specifically contradicts
Consequence preview: What downstream agents or actions depend on this decision
Resolution controls: Select output A / Select output B / Request additional agent run / Abort run
Decision attribution: The human's resolution decision is recorded and attributed in the audit trail

Accessibility

Related Components

HC10 Network Degraded State — may co-occur with conflict; agent failure can produce contradictory partial outputs
HC01 Human-in-the-Loop Controls — the single-agent counterpart for requiring human approval before action
AA01 Audit Trail Widgets — the resolution decision must be recorded as a human governance event
AA03 Aggregate Audit Rollup — conflict events are surfaced in the network-level audit
Pattern 04: Multi-Agent Governance — Stage 02: Conflicting Output Resolution

← Previous Network Degraded State Next → Agent Handoff Receipt