AI Agents Give Contradictory Outputs: Causes and Fixes

Issue: Two or more agents disagree and nothing in the pipeline resolves the conflict, so contradictory answers reach the end user or downstream system.

Frequency: Common

Symptoms

• Conflicting recommendations without arbitration.
• End users or downstream systems receive two mutually exclusive answers (e.g., “approve” and “deny”) with no indication which to trust.
• Disagreement is detected only after the fact by a human, because no automated arbitration step runs before output is finalized.
• The same disagreement recurs for similar inputs because there is no learned or codified tie-breaking rule.

Root Cause There is no designated arbiter agent or deterministic tie-breaking policy defined anywhere in the pipeline, so when two agents optimize genuinely different objectives – growth versus margin, for instance – with no shared reconciliation criteria, their outputs have nowhere to converge. The aggregation step makes this worse by concatenating or passing through whatever each agent produces rather than requiring the outputs to agree before being finalized. Since the agents run in parallel with no communication channel to negotiate or even flag the conflict to one another, neither agent nor the system has any opportunity to notice the contradiction before it reaches the end user.

Example

Agent A (pricing agent) recommends a 15% discount based on customer loyalty tier.
Agent B (margin-protection agent) recommends a 0% discount based on current
inventory cost pressure. Both recommendations are appended to the customer-facing
quote generator as-is; the generated quote reads "recommended discount: 15%" in one
section and "no discount available" in another, with no arbitration step ever
invoked.

Contributing Factors

• No designated arbiter agent or deterministic tie-breaking policy for cross-agent disagreement.
• Agents optimize different, non-overlapping objective functions (e.g., growth vs. margin) with no shared reconciliation criteria.
• Output aggregation pipeline concatenates or passes through all agent outputs rather than requiring convergence before finalizing.
• Agents run in parallel/independently with no communication channel to negotiate or flag conflicting conclusions to each other.

Eval Recipes

Test Cases

Metrics

How to fix it: add an explicit arbitration step that runs before output is finalized instead of surfacing raw disagreement.

Mitigation Strategies

Prevention

1. Implement handoff schema validation with type checking: Define explicit message contracts between agents using JSON Schema or Protocol Buffers. Each handoff includes: required fields, field types, context cardinality, consistency invariants (e.g., ‘account_id must match previous context’). Validation layer rejects malformed handoffs before forwarding. Root cause: Prevents information loss and state inconsistency by catching misalignment at handoff boundaries.

2. Establish distributed consensus checkpoints: Before critical transitions (agent A -> B), compute and store world-model checkpoints as semantic hashes of key state variables. On agent B entry, verify checkpoint matches derived from B’s initial inputs. If mismatch, trigger rollback or human escalation. Root cause: Detects state divergence early, enabling recovery before cascading errors.

3. Implement error isolation with saga pattern: Structure multi-agent workflows as compensating sagas. Each agent’s action has a reverse operation. On error, compensating actions execute in reverse order, restoring system to consistent state. Track saga state in distributed ledger (event log). Root cause: Prevents cascade failures by ensuring partial execution doesn’t corrupt global state.

Detection & Response

1. State consistency verification at handoffs: At each inter-agent message, verify: (1) Handoff schema conforms to contract, (2) Required fields present and non-null, (3) Semantic consistency (e.g., derived context matches explicit assertions). Log mismatches with full message context. Alert on schema violation rate >0.5%.

2. Distributed tracing with invariant checking: Instrument all agent-to-agent calls with trace IDs. Track state variables across spans. Compute invariant violations: e.g., total_balance should equal sum(accounts). Flag spans where invariants break. Correlate with handoff timing to identify failure point.

Architecture Patterns

1. Handoff Contract Engine: Define per-workflow interaction schemas with required fields, optional extensions, and invariant predicates. Codegen produces type-safe message classes. Validation happens pre-send with detailed error reporting (which field failed, why).

2. Saga Pattern with Event Sourcing: All agent actions append to immutable event log. On failure, replay log in reverse (applying compensating actions) to reach consistent state. World-model reconstructed from event log deterministically.

3. Distributed Tracing + Invariant Monitor: OpenTelemetry spans track all inter-agent messages. Background service computes invariants (e.g., sum checks, state graph acyclicity) against live span data. Alert on invariant violation with full context trail.

Key Metrics

Alerts & Escalation

Production Signals

Key Metrics

Alerts

References

• MAST
• Note: Multi-agent system failure taxonomy with system design, inter-agent misalignment, and task verification failures.