🤖 AI

What Should Agents Say? Action-state Communication for Efficient Multi-Agent Systems

This paper introduces PACT, a protocolized action-state communication framework that compresses free-form natural language exchanges in multi-agent systems into compact records, significantly reducing token usage and inference costs while maintaining or improving task performance across various system topologies.

Original authors: Chen Huang, Yuhao Wu, Wenxuan Zhang

Published 2026-06-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Chen Huang, Yuhao Wu, Wenxuan Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Over-Share" Meeting

Imagine a team of detectives (agents) trying to solve a mystery. In current systems, every time one detective finishes their work, they write a massive report and paste it into a shared notebook for the next detective to read.

This report includes:

The final answer.
The clues they found.
But also: Every random thought they had, every dead end they explored, every time they re-read a clue, and a long, rambling explanation of how they thought.

The Issue:
As the investigation goes on, this shared notebook gets huge.

It's expensive: Reading a 100-page notebook costs a lot of money (in this case, "tokens" or computing power).
It's slow: The next detective has to read all that noise before they can get to the new clues.
It's confusing: The important facts get buried under pages of "I thought about this, then I thought about that."

The paper argues that we are currently letting agents "over-share" their private thinking process, which clogs up the system.

The Solution: PACT (The "Executive Summary" Rule)

The authors propose a new rule called PACT (Protocolized Action-state Communication and Transmission).

Think of PACT as a strict rule for the team's shared notebook: "No rambling allowed."

Instead of pasting the whole messy draft, an agent must now translate their work into a compact, three-part "Action-State Record" before passing it on:

ACTION: What did I just do? (e.g., "I searched the database.")
STATE: What evidence did I find? (e.g., "I found a receipt from 1995.")
RESULT: What is the useful output? (e.g., "Here is the receipt image.")

The Analogy:

Old Way: Detective A hands Detective B a 50-page diary of their entire day, including what they had for breakfast and every thought they had while walking to the store.
PACT Way: Detective A hands Detective B a sticky note that says: "I went to the store (Action), found a receipt for a blue car (State), and here is the photo (Result). Now you check the car registry."

What They Found

The researchers tested this idea in two different ways:

Split Evidence: Two agents had to share clues to solve a puzzle.
Assembly Line: Four agents worked in a line (Planner → Critic → Refiner → Solver).

They compared PACT against five other ways of communicating (like sending the full text, just the conclusion, or a short summary).

The Results:

No "One Size Fits All": They found that no single way of talking worked best for every situation. Sometimes sending just the conclusion worked; sometimes sending the full text was okay.
PACT Wins: However, PACT was the most consistent winner. It saved a massive amount of "money" (tokens) while keeping the team just as smart, or even smarter.
- In their tests, PACT reduced the amount of data sent by about 39% on average.
- In real-world coding tools (like OpenHands and SWE-agent), it cut the data usage by nearly 50% without breaking the code.

Why It Works

The paper explains that the next agent in the chain doesn't need to know how the previous agent thought. They only need to know:

What action was taken.
What the current situation (state) is.
What the result is.

By filtering out the "internal monologue" (the private thinking) and only sharing the "public facts," the system becomes faster, cheaper, and less likely to run out of memory.

The Bottom Line

The paper concludes that we shouldn't just let AI agents chat in free-form, unstructured paragraphs. Instead, we should force them to follow a strict "protocol" where they only share the essential updates (Action, State, Result). This turns a chaotic, expensive conversation into an efficient, high-speed relay race.

Technical Summary: PACT for Efficient Multi-Agent Systems

Problem Statement
Multi-agent systems (MAS) built on Large Language Models (LLMs) typically rely on unconstrained natural language for inter-agent communication. While agents collaborate via roles, pipelines, or turn schedules, the content they exchange is often free-form, including full reasoning traces, verbose explanations, and redundant restatements. This approach creates a critical inefficiency: as agents forward long internal reasoning traces, downstream agents must repeatedly process this content as part of the shared dialogue history. This "context compounding" rapidly inflates token usage, consumes shared context windows, and degrades both system performance and inference costs, particularly in multi-turn settings involving modern reasoning models.

Current MAS research has largely focused on orchestration (roles, turn-taking, tool allocation) while treating the inter-agent message content as a secondary, unconstrained variable. The paper argues that the content of messages is a central design lever that directly shapes both performance and cost.

Methodology: Diagnostic Analysis and PACT
The authors first conducted a diagnostic analysis across two distinct MAS topologies—a symmetric Split-Evidence Interaction (Setting A) and a Sequential Pipeline (Setting B)—using models ranging from 8B to 32B parameters. They evaluated five common communication strategies:

Full Content: Forwarding the entire free-form output, including reasoning traces.
Concise Generation: Using native non-thinking modes to produce shorter responses.
Conclusion Only: Passing only the final answer/conclusion.
Brief Summary: Explicitly asking for a short free-form summary.
Artifact Only: Passing only the role-specific artifact (e.g., a plan or critique).

Findings from Analysis:

No single strategy is universally optimal across topologies.
Full Content is consistently expensive and redundant.
Concise Generation and Brief Summary are unstable; they often omit task-relevant information needed by downstream agents.
Conclusion Only works in pipelines (where context is supplied by role order) but fails in interactions where evidence is missing.
Artifact Only captures useful content but lacks the necessary protocol to signal intent and grounding, sometimes leading to increased interaction turns.

The PACT Protocol
Based on the finding that effective messages must preserve action-centered information rather than private reasoning, the authors propose PACT (Protocolized Action-state Communication and Transmission).

Core Mechanism: PACT treats inter-agent communication as a public state-update problem. Before an agent's raw output ( $y_t$ ) is appended to the shared history, it is projected by a sender-side function ( $\Pi_{PACT}$ ) into a compact, structured record.
Message Structure: The public message consists of three fields:
1. ACTION: Explicitly states what the sender did or what the next agent should do (clarifying intent).
2. STATE: Records the evidence, observations, or environment feedback grounding the action (providing context).
3. RESULT: Contains the actionable output (e.g., a refined solution, retrieved fact, or tool result).
Privacy Boundary: PACT strictly separates private computation from public communication. Agents retain their full internal reasoning, self-checking, and planning capabilities; only the resulting action-state record is shared. This prevents intermediate deliberation from compounding in the context window.
Harness-Agnostic: The protocol is orthogonal to MAS design choices (roles, schedules) and can be implemented as a lightweight proxy hook in existing systems without retraining models.

Key Contributions

Systematic Analysis: Demonstrated that no fixed communication policy is universally optimal and identified that inter-agent message content is an underexplored but critical design dimension.
PACT Protocol: Introduced a training-free, harness-agnostic protocol that defines a boundary between private computation and public communication, restricting shared history to compact action-state records.
Empirical Validation: Showed that PACT improves the performance–cost frontier across controlled MAS settings and real-world coding harnesses.

Experimental Results

Controlled MAS Settings:
- Split-Evidence (Setting A): PACT achieved comparable or superior F1 scores while reducing token usage by an average of 38.7% across baselines and model scales compared to unconstrained communication.
- Sequential Pipeline (Setting B): PACT achieved comparable or stronger accuracy while using significantly fewer tokens. For instance, on the Qwen3-32B model, PACT used only ~19–23% of the tokens required by Multi-Agent Debate baselines while maintaining or improving accuracy.
- Ablation: Removing any of the three fields (Action, State, Result) degraded performance and increased token usage, confirming the necessity of the complete action-state handoff.
Real-World Coding Harnesses:
- OpenHands: PACT increased the SWE-bench Verified resolve rate from 19.40% to 23.00% (+18 resolved instances) while reducing tokens-per-resolved by 10.3%.
- SWE-agent: PACT reduced input tokens by 50.4% (from 314.6M to 156.0M) and tokens-per-resolved by 47%, with a negligible impact on the resolve rate (25.6% to 24.2%).

Significance and Claims
The paper claims that disciplining the content of inter-agent messages is an effective mechanism for reducing redundant context growth. PACT demonstrates that agents do not need to share their full reasoning traces to collaborate effectively; instead, sharing a structured, grounded action-state record is sufficient for downstream agents to continue the task.

The authors emphasize that PACT is training-free and harness-agnostic, making it a practical, inference-time intervention for existing agentic applications. The results suggest that the "performance–cost frontier" in MAS can be significantly improved not just by scaling models or changing orchestration, but by strictly defining what information enters the shared communication channel. The paper concludes that leaving inter-agent communication as unconstrained natural language is inefficient and that structured protocols like PACT are necessary for scalable, cost-effective multi-agent systems.