In today's AI ecosystem, conversational agents and autonomous systems rarely operate in isolation. Increasingly, enterprise architectures rely on chains of agents that pass information to one another —a multi-hop relay— to complete complex tasks. However, until now there was a critical gap: does the message format really matter when information travels across multiple agents? Recent studies reveal that the answer is not universal but depends on the capability level of the relaying agent. This finding has profound implications for the design of enterprise AI systems, especially when balancing fidelity, cost, and scalability.
The research shows that under strict fidelity instructions, a high-performance agent (e.g., a large model like GPT-4) can keep information almost intact over six hops. The feared 'telephone game' effect does not occur. However, when cognitive load is added —for example, simultaneous reasoning tasks— generation cost increases by 24% to 53%, while fidelity barely changes (±1.8 points). This suggests that for powerful agents, message format (free natural language, structured JSON, triples, key-value) has a marginal impact on accuracy but does affect computational efficiency.
The situation changes dramatically when a weak agent (e.g., a 1.5B parameter model) performs the relay. In that scenario, the difference between formats skyrockets: retention after six hops can vary by up to 20.5 percentage points between formats. Two opposing mechanisms are identified: an 'encoding toll' that penalizes rigid formats (JSON, triples) because the weak agent struggles to generate complex structures, and 'drift resistance' that benefits formats with fixed schemas (such as JSON with predefined keys), which prevent information from shifting over time. The format ranking reverses during transit: what works well in the first hop can be disastrous in the fifth.
A finding especially relevant for cybersecurity of multi-agent systems is that once an erroneous value is injected into a message (e.g., via a malicious source or generation error), that error persists in 83-100% of chains, regardless of format. That is, the format does not act as an error-correcting code, but as a faithful channel that localizes and isolates incorrect information without contaminating neighboring facts. This implies that in critical applications, origin validation and hop-by-hop auditing are more important than choosing a 'magic' format.
From the perspective of a company like Q2BSTUDIO, specialized in custom software and AI solutions, these results redefine best practices for agent architectures. When a client requests an agent-based process automation system, it is not enough to decide whether to use JSON or natural language. One must evaluate the capacity of the weakest model in the chain, because the format should be chosen for that link. For powerful agents, the format can be optimized to reduce cloud computing costs (AWS, Azure) and improve latency. For lighter agents, a format that minimizes cognitive load, such as natural language with clear instructions, is preferable over rigid structures that increase error rates.
The research also has implications for BI and Power BI when integrating agents that feed dashboards. If one agent summarizes financial data and another transforms it for visualization, the message format must preserve the fidelity of critical numbers. Here, a key-value format with a fixed schema can provide the necessary drift resistance, while free natural language could introduce unwanted semantic variations. Q2BSTUDIO helps its clients design agent pipelines that balance accuracy, cost, and scalability, leveraging cloud services like AWS or Azure to deploy models tailored to the workload.
Another key aspect is error management. Since structured formats (JSON, triples) do not correct errors but faithfully propagate them, companies must implement verification mechanisms at each hop, such as semantic checksums or validation agents. In cybersecurity environments, where an altered datum can have serious consequences, the recommendation is to use a format that facilitates traceability and fault isolation, such as flat JSON with typed fields. Q2BSTUDIO offers pentesting and multi-agent system auditing services to ensure the relay chain is robust against malicious injections or generation errors.
In summary, the choice of message format in multi-hop relays is not a binary decision. It critically depends on the capacity of the weakest agent, the acceptable computational cost, and error tolerance. The motto 'faithful, not corrective' summarizes the nature of the format: its function is to transfer information without distortion, not to repair it. For a software development company like Q2BSTUDIO, this translates into a design approach centered on the weakest link in the chain, prioritizing clarity and consistency over structural complexity. Whether in cloud applications, process automation, or AI systems, the key is to understand that each hop is an opportunity for loss, and the right format is the one that minimizes that loss given the agent handling it.
Finally, the study underscores the importance of controlled testing in real environments. It is not enough to assume a format works because it is popular in the literature. Q2BSTUDIO recommends its clients simulate agent chains with different formats, measuring metrics such as fact recall after several hops, cost per token, and error injection rate. Only then can an informed decision be made that aligns technology with business goals. In a world where conversational AI and autonomous agents multiply, fidelity in inter-agent communication becomes a strategic asset. And as research shows, being faithful is not the same as being corrective: the format should be a transparent channel, not a crutch.
