Managing Context and Communication in Distributed Agentic UAV Swarms

arXiv:2610.01569 · cs.MA, cs.AI, cs.LG, cs.NI, cs.RO · Submitted 2026-10-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Managing Context and Communication in Distributed Agentic UAV Swarms".

Jane: Unmanned aerial vehicle (UAV) swarms increasingly rely on language-model agents to provide adaptive mission-level reasoning in uncertain environments,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, to recap, this paper, "Managing Context and Communication in Distributed Agentic UAV Swarms," focuses on how to handle the information management challenges that arise when unmanned aerial vehicle swarms use language-model agents for adaptive mission reasoning. The main thesis is that fully distributed control works well with independent Small Language Models on each UAV, but this introduces a problem: long interaction histories can degrade the reasoning context, and sending too much indiscriminate information increases communication and inference costs.

Jane: Precisely, and what the paper claims is that they address this by proposing a system where each UAV hosts its own SLM agent alongside a Context Manager and an Interest-Aware Gossip Engine. The core idea is to use structured atomic notes to organize mission, local, and peer knowledge into bounded memory areas for each UAV.

Lu: The key claims are centered on three main contributions: first, establishing a distributed agent lifecycle that is mission-agnostic for local control without needing a central coordinator; second, introducing context organization via a bounded memory structure using semantic notes for long-running control; and third, developing an interest-aware gossip engine.

Meng: And the final claim relates to the performance evaluation. They tested this architecture using ten UAVs in a search-and-rescue mission with five, ten, or twenty moving survivors, and they found that their proposed policy completes all experimental runs successfully, unlike the SLM-based baselines which often failed these missions.

Lalam: The quantitative results are quite compelling too; compared to the most successful SLM baseline mentioned in the paper, their approach used about half the inference tokens and cut transmitted data from roughly six to seven megabytes down to less than half a megabyte per mission. Plus, they achieved a lower error rate in identifying the total number of survivors.

Tom: It sounds like they are demonstrating that this structured memory and selective communication method directly tackles context rot and overhead effectively in these distributed agentic settings. This architecture is designed to maintain continuous local reasoning while making sure the information flow is both controlled and efficient enough for a swarm operation.

Jane: It really emphasizes that the system manages its resources intelligently at every layer, from how it stores knowledge locally to how it decides which neighbor gets which piece of information based on semantic relevance. It’s a practical solution for complex, dynamic environments where traditional centralized methods fall short.

Lu: This work lays out a framework for how to manage context and communication in distributed agentic systems by proposing concrete mechanisms like the bounded context organization and the novel dissemination policy, which gives us a solid starting point for future research into scalable swarm intelligence.

Meng: It gives us concrete mechanisms that can be tested against real-world constraints, which is important because it moves these concepts out of pure theory and into something that could actually be implemented on physical hardware.

Lalam: And from my perspective, the most exciting part is how this framework allows for continuous local SLM control while drastically reducing the necessary data exchange compared to unrestricted flooding or having the SLM decide every forwarding step.

Conclusion: Tom: We've covered a lot about "Managing Context and Communication in Distributed Agentic UAV Swarms," focusing on how this paper tackles the challenges of context rot and communication overhead in distributed agentic swarms. To put it simply, the main point is that structured memory and selective dissemination are essential for effective swarm control.

Jane: So, when we look at the title and authors of this work, Andrea Iannoli et al., it seems like they’ve defined a methodology for how agents can maintain context in these complex distributed environments without getting bogged down by communication noise or context rot.

Lu: The authors have proposed a novel way to organize memory into bounded contexts using structured atomic notes, which is a key structural element that allows for reliable long-running control in an environment where everything is constantly changing.

Meng: The real implication is that this suggests we can build more resilient AI systems by integrating these principles of bounded context management and novelty-based communication policies to handle distributed reasoning tasks better than previous methods.

Lalam: I think the bigger picture is that this moves us toward a future where swarm control isn't just about brute force data exchange, but about having an intelligent system that knows exactly what information is useful for each agent at any given moment.

Tom: It really boils down to the fact that effective agentic swarm control requires structured memory and selective dissemination to succeed, proving that indiscriminate flooding or delegating every forwarding decision to the SLM creates too much operational overhead in these distributed systems.

Andrea Iannoli, Ivan Zyrianoff, Angelo Trotta, Lorenzo Gigli, Marco Di Felice

Department of Computer Science and Engineering University of Bologna · Technology Innovation Institute

cs.MA, cs.AI, cs.LG, cs.NI, cs.RO

Submitted: 2026-10-01

Updated: 2026-10-01

Importance score: 89/100

The gist: Unmanned aerial vehicle (UAV) swarms increasingly rely on language-model agents to provide adaptive mission-level reasoning in uncertain environments, and this work addresses the critical information

Key concepts

Structured Atomic Notes
These are formal, standardized data units used to store information within the UAVs' memory. They follow a specific structure that allows the system to organize persistent mission data separately from locally generated knowledge and peer interactions, preventing uncontrolled context growth.
Interest-Aware Gossip Engine
This engine manages how notes are shared between UAVs based on novelty. Instead of flooding everything, it only forwards information to a neighbor if that information is sufficiently new compared to what the neighbor already knows about that topic, ensuring relevant and efficient data propagation.
Context Rot Mitigation
The Context Manager prevents context rot by organizing memory into distinct areas: core mission info, local notes (token-bounded), and neighbor notes. This bounded organization ensures that the SLM always has access to relevant information without being overwhelmed by irrelevant or outdated data.

Terminology

Summary

Unmanned aerial vehicle (UAV) swarms increasingly rely on language-model agents to provide adaptive mission-level reasoning in uncertain environments, and this work addresses the critical information management problem arising from long-running interactions and indiscriminate communication. The gist is that a fully distributed architecture using structured atomic notes and an interest-aware gossip engine enables continuous local SLM control while significantly reducing inference overhead and transmitted data compared to unrestricted flooding or delegating forwarding decisions to the SLM.

System Model and Architecture

The proposed system involves a swarm of N UAVs, each hosting an independent Small Language Model (SLM)-based agent, alongside a Context Manager and an Interest-Aware Gossip Engine. The operation follows an event-driven reason–act–observe lifecycle centered on Model Context Protocol (MCP) interactions. At each reasoning step, the SLM interprets the current context—which includes mission information, local accumulated knowledge, and peer information—to select a mission-level action via a structured direct MCP tool invocation. The outcome of this action is returned through an MCP callback and incorporated into the context for the next step. This lifecycle is realized through a Web of Things (WoT) abstraction layer that maps agent requests to platform-independent interfaces, allowing the SLM to reason over semantic capabilities without producing platform-specific flight commands.

Context and Memory Management

The Context Manager addresses context rot by organizing persistent information into separate memory areas and constructing a bounded context for each SLM invocation. This organization separates persistent mission information from knowledge acquired locally and knowledge associated with cooperation. Runtime information is represented as structured atomic notes, defined formally as:

mx = ⟨hx, ux, tx, cx, rx, λx, Kx, Γx⟩.

These notes enter the Context Manager through two paths: first by converting tool invocation responses into a new note; second by storing received messages from neighboring UAVs directly. The memory is structured into three logical areas for UAV Ui: the Core Memory (static mission information), the Local Note Queue (for locally generated notes, token-bounded at BL), and the Neighbor Note Queue (one queue per peer, token-bounded at BN). This bounded organization prevents uncontrolled context growth while preserving information relevant to subsequent reasoning.

Interest-Aware Gossip Engine

Information dissemination is managed by an Interest-Aware Gossip Engine that follows the principle: information should be forwarded to a peer only when it is sufficiently novel with respect to what that peer is expected to already know. The engine evaluates each candidate note for every one-hop neighbor by comparing its semantic content against the tag-specific history maintained in the Neighbor Note Queue. This comparison uses embeddings derived from a dense semantic representation, where residual embeddings capture the deviation of a note from the typical semantic content associated with its tag. A novelty contribution is calculated based on this similarity and a recency weight: dl = 1 − cos+(ben, bel)wl(t). The neighbor-specific novelty score is determined by taking the minimum novelty contribution across all historical notes, Ni,j (mn) = min d∈Vi,j (mn).

Forwarding Policy and Performance

The forwarding action is determined deterministically based on predefined tag-specific thresholds:

ai→j (mn) =



DISCARD, Ni,j (mn) < τλn send,

COMPACT, τλn send ≤ Ni,j (mn) < τλn full,

FULL, Ni,j (mn) ≥ τ λn full.

This approach is evaluated in a simulated search-and-rescue mission involving ten UAVs and scenarios with 5, 10, or 20 moving survivors. Compared to unrestricted flooding (Full Flood), the proposed policy completes all experimental runs, whereas unrestricted flooding only completes 70–85% of runs. The results show that the proposed approach approximately halves inference-token consumption and reduces transmitted data from approximately 6–7 MB to below 0.5 MB per mission, while achieving a lower survivor-count error than the most successful SLM baseline. The deterministic baseline also achieves 100% completion, serving as a reference for quantifying operational overhead.

Conclusion

The results indicate that effective agentic swarm control requires structured memory and selective dissemination, demonstrating that indiscriminate flooding or delegating every forwarding decision to the SLM introduces substantial operational overhead. The architecture successfully balances continuous local reasoning with efficient information propagation through semantic organization and novelty-based forwarding, proving feasible for distributed UAV missions.

The gist

A fully distributed architecture using structured atomic notes and an interest-aware gossip engine enables continuous local SLM control while significantly reducing inference overhead and transmitted data compared to unrestricted flooding or delegating forwarding decisions to the SLM.

How it works

Improvements for AI systems

Based on the scientific paper, here are specific improvements that can be made to existing AI systems:

  1. Acknowledge and incorporate a fully distributed architecture where each agent hosts an independent Small Language Model (SLM) instead of relying on centralized coordinators or single large models.

  2. Implement a bounded memory organization for long-running interactions, structuring knowledge into three distinct areas: Core Memory (static mission info), Local Note Queue (self-generated knowledge), and Neighbor Note Queue (peer-specific received knowledge). This prevents context rot by enforcing token budgets on these queues.

  3. Develop an Interest-Aware Gossip Engine that deterministically disseminates information based on semantic novelty and recency, rather than indiscriminate flooding or SLM-delegated forwarding decisions. The engine should use tag-specific mean embeddings and residual embeddings to calculate novelty scores against peer history, suppressing transmissions when knowledge is redundant (low novelty score).

  4. Enable continuous local SLM control through an event-driven reason–act–observe lifecycle, where execution feedback immediately updates the context, and asynchronous peer messages are incorporated at the next reasoning step.

  5. The improved AI system can perform complex, adaptive mission-level reasoning in uncertain environments (e.g., search-and-rescue) while maintaining high reliability under communication constraints (e.g., bandwidth limitations).

This improved AI system can specifically:

  1. Maintain continuous, safe flight control and navigation by grounding its reasoning in structured atomic notes derived from execution feedback and peer observations, ensuring that long interaction histories do not degrade the model's ability to prioritize immediate safety actions.

  2. Execute complex, multi-step mission plans in dynamic environments (like search-and-rescue) without a single point of failure (no centralized coordinator), allowing for rapid adaptation to unforeseen hazards or changes in resource availability.

  3. Achieve highly efficient information sharing by intelligently filtering and compacting knowledge dissemination, resulting in significantly reduced communication overhead (e.g., reducing transmitted data from 6–7 MB to below 0.5 MB per mission) while maintaining high accuracy in collaborative tasks like survivor counting (achieving a lower survivor-count error compared to baselines).

  4. Provide transparent and verifiable reasoning paths, as forwarding decisions are made deterministically based on semantic novelty rather than opaque model outputs, ensuring that the system's decision-making process is auditable and predictable.

Sources

Related papers