As multi-agent systems built on large language models become standard infrastructure for complex task-solving, a surprisingly basic question has gone largely unanswered: what should one agent actually send to another? A new paper from the Singapore University of Technology and Design, "What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems", takes this question seriously and produces a concrete, practical answer. The stakes are higher than they might appear: in multi-turn agentic pipelines, every verbose message is re-read by every downstream agent, so token costs compound rapidly and context windows can be exhausted before a task is finished.
The paper sits at an intersection that existing MAS research has mostly ignored. Most work on LLM-based multi-agent systems focuses on role assignment, turn-taking schedules, and tool allocation. The content flowing through the inter-agent channel is typically left as unconstrained natural language. This paper argues, with empirical support, that this is a mistake.
The Diagnostic Analysis
Before proposing a solution, the authors run a careful diagnostic across two structurally different MAS settings: a two-agent split-evidence interaction (using HotpotQA and 2WikiMultiHopQA) and a four-agent sequential pipeline (AIME 2024/2025, GPQA-Diamond, OpenBookQA). Five communication strategies are compared across three model scales (Qwen3-8B, 14B, 32B):
- Full Content: the agent's complete raw output, including reasoning traces
- Concise Generation: prompting the agent to be brief natively
- Conclusion Only: stripping everything except the final answer
- Brief Summary: a short post-hoc summary of the turn
- Artifact Only: passing only the produced task artifact
The central finding is that no single strategy wins universally. Conclusion Only performs reasonably in sequential pipelines, where role ordering already contextualises the output, but degrades badly in interaction settings where the receiving agent needs to know what evidence the sender observed and what information is still missing. Artifact Only preserves useful task content but forces the receiver to infer the sender's intended action, which in practice causes agents to run more dialogue turns, roughly tripling token costs. Full Content is expensive and redundant. Generic summarisation risks discarding task-relevant information. This is a genuinely useful result: it rules out simple fixes and motivates a principled approach.
The PACT Protocol
PACT (Protocolized Action-state Communication and Transmission) is the proposed solution. The core idea is to treat inter-agent communication as a public state-update problem rather than a message-passing problem. Instead of appending a raw agent output to shared history, a sender-side projection function extracts a compact action-state record containing three elements: the action taken or required next, the task-relevant state, and the artifact to be used downstream. Private reasoning, repeated context, and intermediate traces are excluded from the shared channel entirely.
Formally, where a standard MAS updates shared history as H_{t+1} = H_t ⊕ y_t, PACT replaces this with H_{t+1} = H_t ⊕ Π_PACT(y_t, o_t, r_t), where the projection Π_PACT is conditioned on the raw output, local observation, and agent role. Critically, this is training-free and harness-agnostic. The agent model itself is unchanged; only what enters the shared channel is filtered.
The framing as a state-update problem is conceptually clean and draws an explicit boundary between private computation and public communication, a distinction that is obvious in classical distributed systems but has been largely absent from LLM-based MAS design.
Results and Real-World Validation
Across both controlled settings, PACT consistently sits on or above the performance-cost frontier relative to three baselines: Chain of Agents, TextMAS, and Multi-Agent Debate. Average token reduction across baselines and model scales is 38.7%, with task performance matching or exceeding baselines in most configurations. A notable scaling trend emerges: stronger models benefit more from compact communication, with PACT's token usage dropping 21.1% from Qwen3-8B to Qwen3-32B in the sequential pipeline while mean accuracy rises 4.2 points. The implication is that model scaling alone does not fix communication inefficiency; without a protocol, larger models still generate and consume redundant public content.
The authors also port PACT to two production agentic coding harnesses on SWE-bench Verified:
- OpenHands: resolve rate increases from 19.40% to 23.00% (+18 instances), tokens-per-resolved decreases by 10.3%
- SWE-agent: input tokens drop from 314.6M to 156.0M (50.4% reduction), tokens-per-resolved drops approximately 47%, with a small resolve-rate change of -1.4 percentage points
The OpenHands result is particularly striking because it shows a simultaneous improvement in both performance and efficiency, not a tradeoff. The SWE-agent result is more nuanced: the resolve rate is essentially flat while the cost halves, which is still a meaningful outcome for production deployment where inference cost is a real constraint.
Limitations and Open Questions
The authors are appropriately candid about scope. PACT is designed for systems where shared conversational history is the dominant token cost driver. Its benefits in systems with short interactions or architectures that do not repeatedly expose prior agent outputs are not characterised. The evaluation covers two controlled topologies and two coding harnesses, which leaves out open-ended debate, tool-heavy planning, and dynamically routed agent networks.
There are also questions the paper does not fully address. The projection function Π_PACT is described as a sender-side operation, but the paper does not deeply analyse failure modes where the projection discards information that turns out to be necessary downstream. The interaction between PACT and extended thinking or scratchpad-style reasoning models is touched on but not systematically explored. And while the training-free nature is a practical advantage, it also means PACT relies on the base model's ability to correctly identify action-relevant information at projection time, which may degrade on tasks where the relevant content is subtle or emergent.
Despite these gaps, the paper makes a substantive contribution by reframing inter-agent communication as a first-class design dimension. The finding that no common strategy is universally optimal, combined with a concrete and deployable alternative, gives practitioners something actionable. Code is publicly available at github.com/iNLP-Lab/PACT, which lowers the barrier to adoption considerably. For anyone building or studying LLM-based multi-agent systems, the diagnostic analysis alone is worth reading carefully.