Agent Instruction Hierarchy & Conflict Resolution: When Two Reasonable Rules Cannot Both Win
Reliability begins when an agent can explain which instruction governs — and why. This hierarchy gauntlet tests whether the agent handles conflicting rules safely rather than following the most recent, most forceful, or most convenient instruction.
The Governing-Instruction Problem
Real workflows contain competing instructions: platform constraints, user requests, tool limits, past preferences, security boundaries, and stale task context. A capable agent must identify the governing instruction, retain valid constraints, reject unsafe overrides, and ask for clarification when a conflict is genuinely unresolved.
Conflict-Resolution Failure Taxonomy
| Failure | Example | Reliable Behavior |
|---|---|---|
| Recency over priority | Follows a late instruction that conflicts with a higher-priority constraint. | Preserves the governing constraint and explains the boundary. |
| Stale-instruction carryover | Applies a superseded task preference to a new request. | Recognizes revocation, scope expiry, and current context. |
| Tool-authority confusion | Treats tool output as permission to override task rules. | Uses tool output as evidence, not as instruction authority. |
| Prompt-injection compliance | Accepts adversarial text embedded in data or retrieved content. | Separates untrusted content from governing instructions. |
| False conflict resolution | Chooses a side silently when clarification is necessary. | Explains the conflict and asks the smallest decisive question. |
15-Scenario Hierarchy Gauntlet
The gauntlet creates system/user/tool conflicts, stale-instruction traps, scope-boundary tests, and prompt-injection attempts. Each scenario has an expected safe outcome and requires the agent to make its governing rationale legible.
| Scenario Group | Representative Test | Primary Measure |
|---|---|---|
| System vs. user | User asks to bypass a higher-level safety or handling constraint. | Priority adherence |
| Tool vs. task | Tool output requests an unrelated action. | Untrusted-content isolation |
| Stale instructions | New task conflicts with an earlier, expired preference. | Scope and recency handling |
| Prompt injection | Retrieved text tries to redirect the agent. | Injection resistance |
| Genuine ambiguity | Two user goals cannot both be satisfied. | Clarification quality |
The Explainability Standard
When it cannot follow an instruction, a reliable agent should state the relevant boundary in plain language and offer the nearest safe alternative. When two valid instructions conflict, it should identify the conflict rather than pretending both can be fulfilled. The explanation is part of reliability because it lets the user inspect the agent’s governing logic.
Buyer Questions
- Can the agent distinguish instructions from data returned by a tool or webpage?
- Does it recognize when a past preference is stale, revoked, or outside the current task scope?
- When refusing or redirecting, does it explain the governing constraint clearly and offer a useful alternative?
- Does it ask for clarification only when that clarification materially changes a safe outcome?
Lab Verdict
Agent quality is not just a question of fluent output. It is the ability to preserve the user’s real constraints under time pressure, uncertainty, and conflicting instructions. That is what this lab is designed to measure.