A switching rule for costly conversations
A preprint describes a way for teams of language-model agents to manage communication as they work toward agreement. A harness selects between complete communication and a ring topology, a fixed communication loop, according to the team's collective disagreement. The two modes carry different rates of disagreement contraction and different token costs. The central question is whether an LLM multi-agent reasoning process can reach practical consensus before its computational resources run out.
In the formal model, this is a discrete-time switched system: the team moves through rounds, and the active communication mode can change. A contraction rate is the model's measure of how strongly a mode reduces disagreement. The paper's theorem gives a finite bound on the rounds needed to reach epsilon-consensus, meaning disagreement below a chosen tolerance, and pairs it with a communication-cost formula based on the number of rounds spent in each mode and that mode's token cost.
The result is conditional. The framework assumes synchronous activation, zero communication delay, switching controlled by an external harness, and doubly stochastic weight matrices, a technical condition that keeps the opinion weights balanced. Its budget certificate requires enough starting budget for the predicted communication cost and states that the system then enters the epsilon-consensus safe set with nonnegative budget left. That describes a guarantee inside the stated setup, not a blanket promise about every live AI team.
The illustrative tradeoff
The illustrative numerical experiment used five agents to plan a cloud software architecture. It set the switching threshold at 0.1, the consensus tolerance at 0.03 and the initial token budget at 2,000. The simulation assigned the ring mode a contraction rate of about 0.724 and the complete mode a rate of 0.25.
All three strategies entered the epsilon-consensus safe set, but their routes differed. Adaptive switching finished with final disagreement of 0.0254 after three rounds and used 800 tokens. Complete-only communication reached 0.0188 in two rounds using 1,200 tokens. Ring-only communication reached 0.0274 in six rounds using 600 tokens.
That illustrative run put the strategies in a clear speed-cost order: complete-only was fastest but used the most tokens, ring-only was slowest but used the fewest, and adaptive switching landed between them. The paper's corollary makes the same theoretical comparison, placing adaptive expenditure between ring-only and complete-only bounds and describing adaptive switching as faster than ring-only but cheaper than complete-graph communication.
The tight-budget adaptive run showed where the certificate can fail. With an initial budget of 650 tokens, it spent 600, triggered budget failure at round 2 and remained at final disagreement 0.0752. It therefore failed the budget check before reaching the target tolerance.
Live agents were less predictable
The paper also reports live LangGraph validation for the software-architecture planning task. Each deployment used five specialized agents and tested GPT-5.4-Mini, Gemini 2.5-Flash and Claude Opus-4.8 with adaptive, complete-only and ring-only strategies. The runs used a switching threshold of 0.2, a consensus tolerance of 0.05 and a starting budget of 100,000 tokens. Disagreement was calculated algebraically, while input-token metadata supplied the communication-cost measure.
Under adaptive switching, GPT-5.4-Mini reached consensus in nine rounds with reported expenditure of 41,727 tokens. Gemini 2.5-Flash took 14 rounds and 92,129 tokens, while Claude Opus-4.8 took 11 rounds and 84,944 tokens. GPT's adaptive run was the shortest and had the lowest reported token expenditure among the three tested deployments.
Ring-only also reached consensus within budget, in 17 rounds for GPT, 18 for Gemini and 13 for Claude. It required more rounds than adaptive communication in each corresponding deployment, but remained budget-feasible. Claude's complete-only run triggered budget failure and exited the safe set.
One live result exposed the limits of applying the formal assumptions directly to model behavior. In the GPT complete-only run, disagreement rose after rounds 7 and 9, and the reported contraction rates for those rounds were greater than 1. That behavior shows why the formal theorem and budget certificate must be read as conditional on their assumptions.
A bounded conclusion
The authors interpret the live deployments as demonstrating an efficiency gain from adaptive communication over both alternatives. The reported runs support a narrower conclusion: adaptive switching occupied the predicted middle ground in the illustrative simulation and reached the target for all three named models, while ring-only took more rounds and Claude complete-only failed its budget check. The live observations leave the paper's certificate as a conditional planning tool, rather than an automatic guarantee for arbitrary LLM teams.
The document is labeled arXiv:2608.25099v1, dated Aug. 25, 2026, and says it was submitted on Aug. 17. Its front matter reports support from NSF Grant No. DMS-2315862.
Paper data and sources
Original title: A Control-Theoretic Approach for Resource-Aware Consensus in Multi-Agent AI
Authors: James Flagg, Esteban A. Hernandez-Vargas
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text