Induced Friction Between AI Agents: A Search for Disruptive Solutions

Let AIs compete against each other. AI collaboration may be overrated, or rather, consensus among AIs may kill innovation.

If we're looking for disruptive and innovative AI solutions to today's problems, the biggest obstacle isn't a lack of analytical capacity, but rather that the most likely responses tend to be the most predictable, a product of their average training, and this is what any consultant would say in a first meeting.

To reach another level, we need something different. Instead of AI debate environments, such as collaborative (1,2) or adversarial debates (3,4), whose results can reveal vulnerabilities and convergence traps, as have already been identified (5), a working environment is proposed where multiple AIs run an obstacle relay race where the obstacles are sequential rejections driven by non-judgmental AI arbiter that does not evaluate quality or creativity, but only rejects outputs and measure semantic drift between successive responses. We call this methodology Sequential Stress Optimization (SSO).

The core idea of SSO is simple: instead of promoting debate between agents, which can end in predictable consensus or unproductive, token-consuming loops (5), the proposal is a sequential relay framework with obstacles. Each AI agent receives the previous result from another agent with a single instruction: beat it by divergence. This is a friction scheme built upon successive, non-qualitative rejections, executed across two distinct phases: Exploring and Polishing.

How does it work in practice?

Let's take a real engineering problem: the bottleneck in the submittals review process.

We pose the problem to the first AI agent. Its response is neither corrected nor discussed; it's picked up by the key actor in this scheme: the AI Arbiter.

The AI ​​Arbiter does not generate ideas or judge content qualitatively. Its rule is procedural: in the first phase, it rejects a percentage of the proposals submitted by AIs, let's say 70%, regardless of what each one says.

The Rejection Threshold:

The 70% rejection rate is an arbitrary benchmark set here to illustrate the procedure. In practice, this threshold is a parameter that should be adjusted based on computational cost and solution quality. It can be adjusted dynamically by the user or adaptively by the AI Arbiter (user-selectable option), if the cost outweighs the benefit of optimization (e.g., when the semantic differential decreases). A detailed analysis of the fixed and adaptive rejection rates and their results is beyond the scope of this publication, as validating them for various scenarios requires a significant compute time and token expenditure; therefore, a detailed statistical analysis is reserved for future publication.

Convergence criteria:

The first response (e.g., "hire more reviewers") is discarded. Once this result is obtained, the AI Arbiter passes it on to a second AI with an explicit instruction: "This was already rejected as obvious; find the root flaw and propose something different".

From this succession of responses passed between AI agents, it is pushed out of its statistical comfort zone, allowing potentially innovative solutions to emerge. For example: replacing sequential review with parallel review across disciplines. In successive iterations, increasingly better and more radical proposals are reached. This is no longer optimization; it's a model shift.

When to stop?

Chained rejection cannot continue indefinitely. The first stopping signal occurs when, after reaching the established rejection rate for the first phase, in this case, 70%, the semantic differential between consecutive solutions approaches a minimum. The AI Arbiter takes this signal to mean that the solution space in that direction has been exhausted. If this semantic minimum proves difficult to reach, the AI Arbiter grants an additional 5% margin of rejections threshold for this phase and treats the identified solution as optimal.

Once the first phase, Exploring, concludes, the AI Arbiter activates the second phase: Polishing. Here the goal is to refine the solution from the first phase; the solution obtained is subjected to a new sequence of rejections, now capped between 20% and 30% of everything generated. After that, the AI Arbiter again identifies the minimum semantic dispersion among the results and extracts the optimal solution from the system. This optimal, tentatively disruptive solution is what the AI Arbiter identifies as the final result.

SSO Flow Chart:

Risk and Operational Feasibility Considerations.

Sequential Stress Optimization carries practical risks. Rejecting 70% of the most probable responses as referenced above pushes agents into low-probability territory, where a wrong answer or a hallucination can be argued with full logical consistency. The result can look entirely correct without being accurate. AI Agents may also learn to disguise familiar answers in complex language to pass the Arbiter's filter without changing the underlying idea, so the first phase rejection output will need human follow-up to check substance.

The two failure modes mentioned above, hallucinations and disguised responses, point to the same gap: the Arbiter measures the semantic distance, not the meaning. To compensate for this, a new actor is added to the SSO scheme (see modified flow chart below): the Dynamic Divergence Refiner (DDR) agent. The Dynamic Divergence Refiner (DDR) performs a parallel qualitative check on each sequence. While the Arbiter measures semantic distance, the DDR verifies whether the new response remains within the problem domain and whether its divergence is substantive rather than merely lexical. When the DDR detects disguised repetition or a logically coherent but impractical answer, it vetoes the handoff and restarts the next agent from the last valid state. If drift persists across consecutive rejections, the DDR rewrites the next agent’s input by explicitly naming the repeated concept to exclude and the neglected variable to pursue.

Modified SSO Flow Chart (SSO+DDR):

The Cost Side:

Repeated rejection cycles in SSO imply repeated token generation, which increases latency, token usage, and computational cost. This is even more pronounced in the SSO-DDR scenario, where these costs increase even further. For this reason, SSO is not considered a default method. It is suitable for cases where the value of a better response justifies the additional cost: strategic decisions, not routine tasks. The SSO+DDR scenario would be reserved exclusively for the pursuit of innovation.

Conclusion:

The SSO as a problem-solving approach is not a multi-agent friction system in the traditional sense. It is not based on debate, consensus building, or adversarial interaction between agents It's an obstacle relay race involving multiple AI agents with different analytical approaches. In this race, induced friction arises from the predisposition of an Arbiter to reject the solutions submitted sequentially by each agent. Therein lies the possibility that by promoting difference, something unexpected may emerge, such as a more creative or disruptive solution.

The diversity in agent training provides the necessary variance in the explored sequential solutions. While SSO is architecturally valid with a single agent with different activated modes, its effectiveness decreases with the homogeneity of the agents involved.

On the other hand, the SSO approach identifies two likely failure modes under friction: hallucinations and disguised responses. Their occurrence depends on the complexity of the problem, and to address them, a modified SSO is conceived, with a more complex architecture that includes the Dynamic Divergence Refiner agent (DDR). This agent assists the arbiter by performing qualitative checks that the arbiter does not. The SSO is affected by the computational cost of sequential rejection. The addition of the DDR agent further increases this computational cost, making it significantly more expensive. Therefore, SSO+DDR is recommended for special cases where the goal is outside-the-box solutions.

The SSO and SS+DDR scenarios are theoretical proposals that require validation in terms of both results and costs, the latter being the main constraint. It is estimated that these tests will begin with two or three agents.

 

 References:

 

1. J. Anglen. “AI Agent Debate Systems: How Multi-Agent Collaboration Improves Decision Making”. 2026. RUH.AI.

2. E. Schepis “Democratic Multi-Agent AI: Debate-Based Consensus”. 2025. Medium.

3. J. Mishra. “When Your AI Needs an Enemy”. 2026. Medium

4. Yao et al. “Peacemaker or Troublemaker”. 2025. ArXiv.

5. Zhang et al. “Stop Overvaluing Multi-Agent Debate”, 2025. ArXiv.

Transparency Statement: The author acknowledges the use of Artificial Intelligence (AI) as an assistive tool during the research, data structuring, and content optimization process. The core concept, final review, and critical analysis remain the sole responsibility of the author.

Spanish version: https://ingconcurrente.blogspot.com/2026/06/la-friccion-inducida-entre-agentes-ia.html