PACT: Optimizing Multi-Agent Communication for Efficiency

🔑 Enhanced Key Takeaways

•PACT's development was driven by an analysis revealing that no single fixed communication strategy is universally optimal in multi-agent systems, and that effective inter-agent messages consistently prioritize action-centered information for subsequent agents.
•The protocol establishes a clear distinction between an agent's private computational processes and its public communication, ensuring that only essential, receiver-facing information is transmitted.
•PACT demonstrates significant token usage reductions, averaging 38.7% across various baselines and model scales, while either maintaining or enhancing overall task performance.
•In practical applications, PACT substantially improves the resolve rate of OpenHands while reducing tokens-per-resolved by 10%, and achieves resolve-neutral performance on SWE-agent with a 47% reduction in input tokens.
•When compared to 'Multi-Agent Debate' communication strategies, PACT achieves comparable or superior performance using only 19%–23% of the tokens, indicating that compact, action-related artifacts are more beneficial than extended deliberation in certain multi-agent pipelines.

📊 Competitor Analysis▸ Show

Feature	PACT (Protocolized Action-state Communication and Transmission)	PAX Protocol (Parallel Agent eXchange)	Unconstrained Natural Language	MCP (Model Context Protocol)	ACP (Agent Communication Protocol - IBM/Linux Foundation)	A2A (Agent-to-Agent Protocol - Google)
Primary Goal	Optimize multi-agent communication for efficiency, reduce token usage, maintain/improve task performance.	Reduce token waste in multi-agent communication.	General human-like communication.	Standardize model access to external tools and data.	Standardize communication between diverse AI agents for interoperability.	Enable secure, peer-to-peer collaboration and task delegation between autonomous agents.
Token Efficiency Focus	High: Projects raw outputs into compact action-state records (ACTION, STATE, RESULT). Reduces tokens by 38.7% average.	High: Enforces "caveman-style" outputs, no pleasantries, fragments, structured results. Cuts ~60-70% token waste.	Low: Free-form, verbose, rapidly inflates token usage.	Moderate: Focus on structured context, not explicit token reduction for inter-agent messages.	Moderate: Focus on structured messages for interoperability, not explicit token reduction.	Moderate: Focus on structured messages for collaboration, not explicit token reduction.
Communication Structure	Structured action-state records (ACTION, STATE, RESULT).	Structured format (e.g., ACTION, STATUS, KEY_DATA, BLOCKERS, NEXT).	Unconstrained natural language.	JSON-RPC protocol for tool/data access.	REST/HTTP native, standardized message formats.	JSON over HTTP, "Agent Cards" for capabilities.
Performance Impact	Comparable or stronger task performance with substantially fewer tokens. Lifts OpenHands resolve rate, resolve-neutral on SWE-agent while halving tokens.	Improves efficiency by reducing token cost.	Can negatively affect system performance and inference cost due to context window inflation.	Improves multi-agent system performance by standardizing context sharing.	Enables seamless interoperability, improving scalability.	Facilitates dynamic agent ecosystems, improving collaboration.
Open Source/Availability	Code publicly available on GitHub.	Part of an open-sourced multi-agent starter kit.	N/A	Widely adopted, several vendors delivering MCP servers.	Open-source, hosted by Linux Foundation.	Open standard, co-developed with partners, hosted by Linux Foundation.

🛠️ Technical Deep Dive

PACT is designed as a minimal and harness-agnostic communication protocol for multi-agent systems.
It operates by projecting each non-terminal agent's raw output into a compact public action-state message.
The protocol is engineered to retain only the receiver-facing information necessary for the continuation of tasks, excluding intermediate process-level content.
Inter-agent communication is conceptualized as a public state-update problem.
Each public inter-agent message is structured as an action-state record, comprising three distinct fields:
- ACTION: Explicitly states the action performed by the sender or the action required from the next agent, clarifying handoffs.
- STATE: Records grounding information such as evidence, observations, environment feedback, or results from tool usage.
- RESULT: Contains the direct output of the action, which could be a retrieved fact, a refined solution, or the outcome of a tool call.
PACT is a training-free protocol, meaning it does not require additional model training.
The implementation code for PACT is publicly available on GitHub.

🔮 Future ImplicationsAI analysis grounded in cited sources

Widespread adoption of PACT could significantly lower the operational costs of large-scale multi-agent systems.

By drastically reducing token consumption without sacrificing performance, PACT directly addresses a major cost driver in LLM-based multi-agent systems.

PACT's structured communication approach will foster greater interoperability and robustness in multi-agent systems.

Standardized, compact action-state records reduce ambiguity and facilitate clearer handoffs between diverse agents, making systems more reliable and easier to integrate.

The principles behind PACT could influence the design of future LLMs to inherently generate more concise and action-oriented outputs for agentic use cases.

If structured communication proves consistently superior for agent collaboration, LLM fine-tuning or architectural modifications might prioritize generating PACT-like outputs directly.

⏳ Timeline

2026-06

PACT (Protocolized Action-state Communication and Transmission) paper published on arXiv.

PACT: Optimizing Multi-Agent Communication for Efficiency

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (9)

👉Related Updates