acceptodds
Under review as a conference paper at ICLR 2027

Task-Aware Communication for Distributed LLM Agent Systems

Abstract

Multi-agent systems (MAS) powered by large language models increasingly coordinate reasoning, tool use, and data processing across distributed environments, where task completion time depends largely on the quality of network connectivity. Existing MAS frameworks focus on managing task workflow and resources on the application level, overlooking the opportunities for performance optimization in the underlying network substrate. Building on the emerging path-aware networking paradigm, we propose *FlowAhead*, a framework that leverages task information and network measurements to choose optimal paths for inter-agent data transfers. Our evaluation shows that *FlowAhead* reduces the median task completion time of a three-agent video-reasoning workflow by 34.5% relative to default BGP routing, and by 51.5% without site-access limits, demonstrating a promising path toward improving efficiency for distributed MAS.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Distributed Multi-Agent Systems, Task-Aware Communication Planning, Network Resource Allocation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_yPinOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FlowAhead, a task-aware network path selection framework for distributed LLM-based multi-agent systems. It uses currently revealed workflow dependencies, object-transfer information, and network measurements to select paths intended to reduce downstream agents' input-waiting time.

The authors evaluate FlowAhead using two multi-agent workloads: long-video question answering (A4VL) and multi-agent debate. Experiments are conducted in an emulated network derived from the GEANT topology, comparing FlowAhead against default BGP routing and random path selection under different background loads, access-bandwidth limits, and team sizes. The paper reports substantial task-completion-time improvements for video-heavy workloads, including a median reduction of 34.5% for three-agent A4VL at 25% background load with a 100 Mbit/s site-access limit. Improvements for text-based debate are considerably smaller.

Critical strengths
  1. A practically relevant systems optimization problem.

The paper identifies an important performance consideration for geographically distributed agent workflows: intermediate data transfers may delay downstream execution, particularly when workloads involve large artifacts. Using workflow dependencies to guide network resource allocation is a potentially useful direction.

  1. Controlled measurements of end-to-end performance.

The evaluation measures both network transfer time and end-to-end task completion time. Replaying identical model calls across routing policies helps isolate the effects of network path selection. The comparison between video-heavy and text-heavy workflows also provides useful evidence that network optimizations have different impacts depending on the communication characteristics of a workload.

Critical weaknesses
  1. The evaluation does not isolate the contribution of task awareness from ordinary network-aware path selection.

This is my main concern. The central contribution of FlowAhead is its use of multi-agent workflow dependencies to improve path selection. However, the experiments primarily compare against modeled default BGP routing and random path selection. Neither is a sufficiently strong control for determining the incremental benefit of task information.

A network-aware baseline that selects paths using the same available bandwidth, latency measurements, and candidate paths, but without workflow dependencies, is essential. In fact, Section 3.2 already describes a network-only policy used to initialize FlowAhead's search, making this a particularly natural ablation.

This distinction is important because the substantial reported speedups may arise primarily from avoiding congested network paths. For example, Appendix C.1 describes a major video-transfer route where the modeled default path provides approximately 20 Mbit/s, while the selected alternative provides 100 Mbit/s under the access limit. A task-agnostic network-aware selector could potentially obtain much of this benefit.

There is also an ambiguity between the algorithmic description and the evaluation protocol. Section 3.2 and Algorithm 1 describe joint path assignment based on predicted consumer input-readiness time, while Section 4.1 describes FlowAhead as selecting the path with the smallest predicted transfer time based on RTT and available bandwidth. The paper should clarify exactly which selection procedure produced the reported experimental results.

Without a matched network-aware baseline and a clear decomposition of the proposed algorithm's contribution, the current experiments demonstrate the benefit of improved network path selection, but do not convincingly establish the additional benefit of task-aware optimization.

  1. Limited evidence of robustness and generalizability beyond the particular emulated network conditions.

The evaluation uses a single emulated network topology and two workflow types, with three- and six-agent configurations. The video evaluation contains only 20 tasks, while the debate workload uses 100 GSM8K questions. These experiments provide an initial proof of concept, but they offer limited coverage of the diverse dependency patterns and dynamic behaviors of distributed agent systems.

The network model also relies on several constructed assumptions. Background traffic is represented through effective link rates rather than transmitted background packets, saturated links receive a fixed additional delay, and packet loss and retransmission effects are omitted. Moreover, FlowAhead receives current available-bandwidth and delay information from the emulator, whereas a deployed system would need to estimate these quantities.

These choices are reasonable for an initial controlled experiment, but the robustness of the reported gains to more realistic network conditions remains insufficiently established. In particular, the results appear sensitive to the presence of a strongly congested default route and the availability of an alternative path with substantially better network characteristics.

Finally, the proposed online task-aware comparison falls back to a network-only assignment when relevant consumer dependencies or object availability information is unresolved. The paper does not adequately quantify how frequently its task-aware optimization is actually applicable during the evaluated workflows.

Stronger sensitivity analyses, realistic measurement uncertainty, and more diverse workflow dependency structures would help establish whether the method generalizes beyond the selected experimental setup.

Questions for authors

Q1. Isolating the benefit of task awareness.

Can the authors compare FlowAhead against the network-only path-selection policy already described in Section 3.2, using identical candidate paths and network measurements? How much additional task-completion-time improvement is obtained specifically from workflow dependency information? Please also clarify whether the evaluated implementation uses the joint input-readiness objective in Equation 7 and Algorithm 1 or the minimum predicted transfer-time rule described in Section 4.1. A controlled ablation isolating task awareness would materially affect my assessment.

Q2. Applicability and robustness of the online policy.

What fraction of path-selection decisions actually use task-aware optimization, and how often does FlowAhead fall back to the network-only policy because of unresolved dependencies or unavailable estimates? How sensitive are the improvements to bandwidth-estimation errors, measurement delays, and the assumed shared-bottleneck model? Could the authors report the runtime overhead of the online path-selection procedure?

Q3. Generalization beyond the evaluated network conditions.

Can the authors provide sensitivity experiments varying the constructed 20 Mbit/s minimum effective-rate parameter, the fixed 50 ms congestion penalty, and the background-traffic distribution? How do the results change under alternative network topologies or less strongly congested default routes? Additional experiments with more diverse, dynamically evolving workflow dependency graphs would also help establish whether the reported benefits generalize to realistic distributed agent systems.

Reviewer_kVDeOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies network path selection for geo-distributed LLM multi-agent systems, where large inter-agent transfers can dominate task completion time. It models execution as an online task graph and proposes FlowAhead, a sender-side middleware on top of the SCION path-aware architecture. Compared with fewest-hop BGP routing and random selection, FlowAhead cuts the median TCT of 3-agent A4VL by 34.5%.

Critical strengths

[S1] Timely problem with a clean online formulation.

Using the network path of inter-agent handoffs as a lever for MAS latency is rarely discussed in agent-systems work. The worked example shows clearly why the same set of transfer times can lead to different readiness, which is the paper’s main conceptual value and a reasonable starting point for follow-up work.

[S2] Transparent measurement.

Replaying identical model calls under every policy isolates the network effect. The paper reports per-task ratios, tail latency and a network/compute breakdown, and openly states assumptions such as oracle bandwidth. This makes the workload-level finding credible.

Critical weaknesses

[W1] The experiments do not test the paper’s main idea, task awareness.

Both baselines ignore network state, and Sec. 4.1 only describes FlowAhead’s network-layer rule: pick the path with the smallest RTT plus size over bandwidth, using oracle bandwidth from the emulator. The paper never reports how often the task-aware search in Algorithm 1 was feasible or changed this choice. So the gains can be fully explained by network-aware path selection, and the paper’s main contribution is not supported by the experiments.

[W2] The headline numbers are mostly set by hand-picked emulation parameters.

Background traffic is never sent. Congestion is modeled as a fixed 20 Ms floor plus 50 ms per saturated link, with no loss, and no sensitivity analysis is given.

[W3] The paper is more a system work.

The paper is more like a computer network / system work, and may not be suitable for ICLR.

Questions for authors

[Q1] Was the task-aware step actually active in the experiments?

How often was the comparison in Algorithm 1 feasible, and how often did it change the network-only assignment? Given the strict fallback rule and workloads dominated by single bulk transfers, I suspect it rarely fires.

[Q2] How robust are the gains to the network model?

Do the conclusions hold in these cases: the 20 Ms floor and 50 ms penalty are varied, background traffic is actually transmitted, available bandwidth comes from the Eq. 5 adapter instead of the emulator?

[Q3] How do the emulated paths map to SCION?

Candidate paths differ at the PoP level inside GEANT, but SCION exposes AS-level paths, and App. A.2 says endpoints do not control intra-AS routes. Is each PoP treated as an AS, and would this path diversity exist in a real deployment such as SCIERA?

[Q4] What is new relative to ASGE-RR?

The introduction says no prior system uses task information to select network paths for agentic communication. Yet Sec. 5 describes ASGE-RR as selecting request/response paths for agent calls based on predicted workflow continuations. Please clarify the difference and compare with previous methods.

Reviewer_d1pxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FlowAhead, a task-aware communication framework for distributed LLM-agent systems. The framework leverages task information and network measurements to optimize communication paths for inter-agent data transfers. The authors evaluate FlowAhead on two multi-agent workflow benchmarks. Experimental results show that FlowAhead reduces task completion time compared with existing baselines.

Critical strengths
  1. The paper connects task dependencies with network-level path selection to address the performance problem in distributed LLM-agent systems. Unlike conventional path selection based only on network metrics, the proposed approach considers when transferred objects are actually needed by downstream agents. This is a meaningful direction for improving end-to-end efficiency in distributed agent collaboration.

  2. The evaluation considers different settings. The reported results show substantial reductions in task completion time, particularly for three-agent video-reasoning workflow. These results demonstrate the practical potential of optimizing network communication for distributed multi-agent applications.

Critical weaknesses
  1. The main contribution of FlowAhead is to incorporate task dependencies into network path selection. However, the experiments only compare FlowAhead against default BGP routing and random path selection. Neither baseline represents a network-aware policy that selects paths based on available bandwidth, latency, or estimated transfer time.

  2. Although FlowAhead is motivated by dependency-aware path assignment and shared-bottleneck contention, the evaluation does not directly quantify how often task information changes the selected paths or improves decisions beyond network-only optimization. The experiments cover only two application workflows, with three or six agents, and use an emulated network. Furthermore, FlowAhead seems to use mean input-readiness time rather than directly optimizing end-to-end task completion time. These limitations make it difficult to assess whether the proposed task-aware mechanism generalizes beyond the evaluated settings.

Questions for authors
  1. How much of FlowAhead's improvement comes specifically from task-aware optimization rather than network-aware path selection? Can the authors provide results or existing measurements comparing FlowAhead against a network-only policy that selects the path with the shortest estimated transfer time using the same network information?

  2. The proposed algorithm seems to minimize the average input-waiting time of affected consumers rather than the overall end-to-end task completion time. Could the authors clarify under what conditions these objectives are aligned?

  3. How sensitive are the reported improvements to the simplified congestion model? Do the authors have evidence that the relative performance gains remain stable under different parameter settings or real network environments?

  4. FlowAhead relies on accurate network-state information to predict transfer completion times and optimize path assignments. However, the evaluation provides FlowAhead with the current available bandwidth and delay directly from the network emulation. The authors acknowledge in Appendix C that a deployed endpoint would need to estimate available bandwidth. This raises an important concern about the practical feasibility and robustness of the proposed approach. The paper does not sufficiently evaluate how measurement errors or network delay affect its performance.

Reviewer_w1vCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FlowAhead, a middleware layer that chooses network paths for inter-agent data transfers in distributed LLM multi-agent systems, on top of the SCION path-aware architecture. At each decision point it starts from a network-only assignment, which predicts transfer time as RTT plus size over available bandwidth. It then runs a local search over one- and two-object path changes. The search accepts a change only if it lowers the mean remaining input-waiting time of known consumers, with shared bottlenecks modelled by equal throughput sharing.The evaluation is an emulation in Mininet with tc, on a 22-PoP GEANT-derived topology with SNDlib background demand. It replays recorded model calls for two workloads: A4VL long-video QA (20 tasks, videos of 85–914 MB) and GSM8K multi-agent debate (100 questions), each with 3 and 6 agents. FlowAhead is compared with fewest-hop BGP and random path selection. The headline results are median reductions in task completion time (TCT) of 34.5% for three-agent A4VL at 25% background load, and 51.5% in an idealised setting without site-access limits. Debate gains are about 6%, and gains at 0–10% load are about zero.

Critical strengths
  • The problem is interesting and timely. It joins agent workflow structure with path-aware networking, and I know of no direct prior treatment.
  • The paired-replay protocol is a good design. Model calls are recorded once and replayed under every policy, so differences isolate path choice. The paper gives absolute and ratio distributions and per-task timelines (Figs. 8–12).
  • The decomposition of TCT into network time and agent time (Fig. 7) explains clearly why large network savings give very different TCT gains for A4VL and Debate.
  • The reporting is candid in places. The authors say the policy gets true bandwidth and delay from the emulator (“favors FlowAhead”). They call the load levels “constructed pressure scenarios” and say future work should isolate the value of task information.
  • The numbers are internally consistent. I reproduced the 56.4% and 96.2% network-time reductions, the 35.9% and 23.5% P90 reductions, and the 1,920 paired comparisons (1,413 faster plus 507 equal).
Critical weaknesses
  • The central claim, that task awareness helps, is not tested. The policy begins from a network-only assignment, but that policy is never reported as a baseline. The Conclusion itself says future work “should isolate the incremental value of task information against a network-aware baseline with matched path and observation capabilities”.
  • The emulation favours the proposal. Background demand is routed along the default shortest paths, the same routes BGP uses. The baseline’s route is therefore the congested one by construction. Gains vanish at 0% load and appear when the BGP route is loaded.
  • The baselines are weak. BGP is modelled as fewest-hop, and random selection is a straw man (it is often worse than BGP). Missing are a network-only SCION or widest-path policy, multipath or coflow-style schedulers (the paper cites Aalo and Siphon), and application-level fixes for 400–900 MB payloads such as compression, streaming or pipelined processing, and colocation. An oracle or exhaustive-search upper bound would show how close the local search gets.
Questions for authors

What are the results for the network-only SCION policy (π_net) alone, and for FlowAhead without join awareness or with oracle information? How often does the task-aware stage change the network-only assignment, and what share of the TCT gain comes from those changes? Can you randomise or vary agent placements and topologies, and add workloads with concurrent branches and joins?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.