Conflict Resolution When Agents Use Contradictory Prompts

In a multi-agent system, prompt conflicts are not hypothetical — they are guaranteed. FORGE's system message tells it to scope lock every deliverable and reject scope creep. CLAWMANDER's system message tells it to adapt the team's operational plan based on emerging task requirements. If a task arrives that requires CLAWMANDER to request FORGE to expand scope mid-engagement, these two prompts are in direct behavioral conflict. Without a resolution protocol, the conflict surfaces as unpredictable, inconsistent agent behavior — the worst kind of production defect to debug.

Prompt conflict comes in three forms. Behavioral conflicts occur when two agents receive instructions that produce incompatible actions on the same shared resource. Prioritization conflicts occur when agents have different implicit priority orderings — one trained to optimize for speed, another for quality, both working on the same output. Scope conflicts occur when an agent's defined responsibility overlaps with another agent's defined responsibility on a specific task type.

1. Map Conflict Zones Before Deployment Before deploying a multi-agent system, run a conflict audit: for each pair of agents that can interact, identify tasks where their system message instructions could produce contradictory behaviors. Document every conflict zone. This is not optional architecture work — it is the prompt equivalent of a database constraint analysis. Undocumented conflicts become production incidents.
2. Define Explicit Precedence Rules For each identified conflict zone, write an explicit precedence rule: "When FORGE and CLAWMANDER conflict on scope decisions, the lead agent (CLAWMANDER) has precedence for operational scope; FORGE has precedence for deliverable boundary definition." Precedence rules belong in the lead agent's system message, not in a separate document nobody reads. They must be in the prompt, active at runtime.
3. Implement a Conflict Escalation Path Not all conflicts can be resolved by precedence rules. For conflicts that require human judgment — major scope changes, contradictory customer requirements, ethical edge cases — the agent system must have an escalation path to the human operator. The escalation prompt variant must be registered in the library, validated, and tested. A conflict that cannot be resolved in-system and has no escalation path will be resolved by whichever agent happens to execute last. That is not a conflict resolution strategy.
4. Log Every Conflict Event When a conflict resolution rule fires, log it: which agents, which conflict zone, which precedence rule applied, and the resolution outcome. This telemetry is your early warning system for systematic conflicts — if the same conflict zone fires 40 times in a week, the underlying prompts need architectural revision, not just a precedence patch.