Semantic Blast Radius: How Errors Propagate Through Multi-Agent Systems

Consider this: a five-agent system engineered to investigate whether a production application can safely migrate to a new database version is deployed. The agents each have their own responsibilities and have a coherent architecture for optimal context utilization and specialized tracking. One agent checks compatibility. Another agent reviews dependencies. A third looks at historical incidents; the fourth one evaluates the migration plan. The final agent takes all those findings and produces a recommendation. The system runs on the production space, but then the compatibility agent makes a mistake by concluding that the new database version supports a critical dependency when it does not. The impact of this error is not contained to the single agent; the dependency agent uses this outcome and posits that there are no compatibility issues. The migration agent and risk agents proceed under the same assumption and lower their assessment of risk. Finally, the decision agent sees several agents arriving at compatible conclusions. Four agents appear to agree. The system approves the migration.We tend to evaluate agent reliability at the agent level: Did the agent select the correct tool? Did it retrieve the right information? Was its reasoning correct?These are important in single-agent systems, but in modern implementations where multiple agents parallelly investigate the depth of a problem space, many failures emerge from how agents interact, not simply from an incapable underlying model. Once an output becomes another agent's input, we need to evaluate cross-agent interactions with a low-trust view.The Semantic Blast RadiusSemantic Blast Radius: the portion of a multi-agent system whose reasoning or decisions become materially influenced by one incorrect or unreliable piece of information.This is an architectural framing for a problem that recent multi-agent systems face frequently. A fundamental principle in infrastructure design is understanding and scoping the blast radius of impact. It describes how much of a system can be affected by a failure. We deliberately create boundaries and safeguards so local failures stay local. Microservices and interactive cloud architectures have matured in this aspect to contain failures while balancing service reliability. Multi-agent systems need a similar way of thinking about information; we need to qualify probabilistic outcomes rather than treat them as infallible. What Causes the Poison to Spread1. Perfect Delivery Can Make Bad Information More ReliableDistributed systems that are deterministic employ many tools for reliability i.e retries, idempotency, checkpoints, circuit breakers, event-driven communication, etc. These concepts are absorbed in modern AI architectures for agent reliability. A multi-agent system adds another layer to this equation: the evaluation of whether a piece of information deserves to be delivered. A message can be delivered, parsed, processed, and semantically understood perfectly while still being factually wrong. The infrastructure would have reliably delivered an unreliable belief.Reliability layerCore questionExample failureTransportDid the message arrive intact?Timeout, loss, duplicate deliveryExecutionDid the receiving agent complete the task?Tool or workflow failureSemanticIs the communicated claim correct and properly interpreted?Hallucinated or distorted claimPropagationShould this claim influence more agents?Unverified claim spreads network-wide2. Every Communication Edge Is Also a Failure PathA common response to improve multi-agent co-ordination is to increase communication. By allowing agents to share more context, they can cross-validate each other’s work and draw different investigation paths to corroborate information. Agents can debate and inspect conclusions.However, this is a double-edged sword. While fully connected systems boost the opportunity to exchange good information, it also allows bad assumptions to spread further. Every communication edge is both a collaboration path and a potential failure-propagation path. Recent research on information propagation in LLM multi-agent networks suggests that maximum connectivity is not necessarily optimal. Moderately sparse topologies can preserve useful information diffusion while suppressing erroneous propagation.The reliability problem is therefore not how to maximize communication. It is how to maximize useful information diffusion without maximizing erroneous information propagation.3. Repetition Is Not CorroborationAgents may present a consensus on a particular decision or fact however, this does not necessarily equate to the verity of the information they present. These systems are susceptible to group-conformity effects where agents can move toward numerically dominant groups or more influential agents. Adding agents therefore does not automatically give us independent reasoning. Sometimes it gives the original mistake more places to echo. A stronger measure is to evaluate the deterministic source of information, if agents have independent investigation paths or study different sources of truth that they can then corroborate, it is a stronger measurement than a simple vote. Consensus should care about where information originated, not simply how many agents eventually repeated it.Figure 2. Three agreeing agents do not equal three independent sources.Qualifying the Semantic TraceSemantic analyses can alleviate the pressure on systems to depend on individual agent sources for ground truths. To qualify semantic propagation it’s prudent to look beyond agent observability and emphasize properties such as:Propagation depth - How many agent-to-agent hops has the information traveled?Propagation width - How many agents have consumed it?Source diversity - How many genuinely independent sources support it?Downstream influence - How many later conclusions depend on it?Verification distance - How far did it propagate before independent validation?Instead of monitoring only an agent graph, we begin monitoring an information graph. The agent graph tells us who talked to whom. The information graph tells us how a belief became a decision.MetricWhat it capturesWhy it mattersPropagation depthNumber of agent-to-agent hopsDeep chains can obscure the original sourcePropagation widthNumber of agents exposedApproximates potential semantic blast radiusSource diversityIndependent evidence originsSeparates corroboration from repetitionDownstream influenceDecisions derived from the claimShows the operational consequence of an errorVerification distanceHops before independent checkingHighlights how long uncertainty travels uncheckedInfrastructure reliability rarely assumes every component should trust every other component indefinitely. In modern distributed systems, we create isolation boundaries. Multi-agent architectures may eventually need an equivalent for information.Figure 3. A reliability layer can constrain propagation based on evidence, provenance, verification, risk, and TTL.Figure 3. A reliability layer can constrain propagation based on evidence, provenance, verification, risk, and TTL.Inside a domain, agents might exchange tentative hypotheses freely. Crossing into another domain could require stronger reasoning or independent verification. A brainstorming agent saying, "I suspect database contention," should not have the same propagation rights as a diagnostic agent saying, "Query telemetry confirms lock contention."The Goal Isn't to Stop Information From SpreadingIf we aggressively contain information, we destroy the reason for building a multi-agent system. Agents need collaboration, it allows them to specialize in dedicated tasks and leverage the power of their expertise which simply cannot be reproduced with a single agent. The objective isn't zero semantic blast radius but rather it should be controlled semantic blast radius. As multi-agent architectures become larger and more connected, the most reliable system may not be the one where every agent can talk to every other agent. It may be the one that knows exactly how far any one agent is allowed to be wrong. The more workflows are automated, there must be an emphasis on the level of trust within the system to produce reliable results. A hallucination inside one agent is a model-quality problem. A hallucination that propagates through an agent network is a systems-reliability problem.References and Further ReadingBerkeley MAST: Multi-agent system failure taxonomy and execution-trace analysis. EMNLP 2025: Research on correct and erroneous information propagation across LLM multi-agent network topologies. ACL 2025: Research on group conformity in LLM multi-agent systems. AgentPrune - ICLR 2025: Communication pruning for multi-agent systems. AAAI 2026: Byzantine fault-tolerance perspective on multi-agent reliability.