1 Introduction↩︎

Modern organizations run on teams: people plan, divide, and verify work under deadlines while juggling synchronous meetings, asynchronous chats/emails, and limited attention. Extending this paradigm to AI, coordinating teams of LLM agents to execute complex, long-horizon tasks (e.g., software development) remains a grand challenge [1]–[3]. Effective collaboration hinges on communication: too much creates coordination overhead; too little yields misalignment and rework.

Recent multi-agent frameworks demonstrate that structured dialogues and role specialization can transform general purpose LLMs into cooperative problem solvers across software engineering, data analysis, and decision support [2]–[5]. Beyond mere tool use, conversation itself decomposes tasks, aligns partial knowledge, and arbitrates proposals, often outperforming single agent pipelines on complex problems [6]–[9]. Yet many systems still schedule or trigger communication via fixed or purely reactive heuristics that ignore the evolving trade-off between coordination cost and task progress [4], [10].

Meanwhile, empirical studies of software teams warn that coordination costs (especially meetings), can dominate delivery time in modern distributed work, with a post-2020 shift toward heavier synchronous load and changing patterns across chat, email, and meetings [11], [12]. These observations motivate treating communication as a first class resource to be scheduled, routed, and evaluated rather than as a byproduct of agent reasoning, thus leading to a central open question: how can agents learn to communicate strategically to maximize collaborative benefit while minimizing coordination costs?

In this work, we address the above question by casting multi-agent collaboration as cost communication and optimizing it accordingly. Our contributions are:

• We introduce Communication to Completion (C2C). A scalable framework models when, whom, and how to communicate as part of execution and instantiate it with a Sequential Action Framework (SAF) that synchronizes agent actions into discrete timesteps for deterministic, reproducible collaboration.

• We define Alignment Factor (AF), a metric that quantifies each agent’s task understanding and directly modulates work efficiency. Communications update AF, and AF in turn scales work rate, linking conversation to progress and yielding interpretable traces.

• We demonstrate through comprehensive experiments spanning complexities and team sizes, C2C lowers completion time and improves work efficiency compared to baselines.

2 Related Work↩︎

2.0.0.1 Multi-Agent Collaboration Frameworks

LLM agents are increasingly organized as teams rather than monoliths. Social simulation work demonstrates emergent, human-like collective behavior (e.g., the “Smallville” agents of [1] and the recent large scale AgentSociety simulation [13]). For goal oriented collaboration—particularly software development—frameworks such as MetaGPT [2] and ChatDev [3] impose predefined roles and SOPs that mirror corporate workflows. AutoGen [4] offers a flexible substrate for composing conversable agents, while AgentVerse [10] and OpenAgents [14] provide extensible environments for building and deploying general purpose agents. Outside pure research prototypes, platforms like OpenHands operationalize multi/single agent development loops in realistic toolchains [15]. In parallel, role-playing and deliberation structures further organize collaboration within conversations [5], [8], [9], and new evaluations emphasize interactive, multi-environment testing and repository-level search and planning [16], [17]. Our work differs in elevating communication from a procedural step to an optimizable resource, explicitly linking when/what to communicate with collaborative progress via an alignment driven mechanism.

2.0.0.2 Communication in Multi-Agent Systems

Communication has long been central in multi-agent learning. Classic MARL studies learned protocols end-to-end under partial observability [18], [19]. Recent advances focus on efficient or sparse communication to reduce overhead, e.g., decentralized scheduling and model-based message estimation [20]–[23]. In LLM agent settings, natural-language messaging carries high-level, semantic content, and methods increasingly structure agent talk through reasoning–acting loops (ReAct) [24], self improvement with textual feedback (Reflexion) [25], and iterative self-refinement [26]. Multi-agent debate further uses conversational arbitration to improve correctness and planning [27]. Yet many systems still trigger communication on fixed or purely reactive heuristics. We instead propose an alignment driven policy that decides when, whom and how to communicate based on a measurable, task grounded alignment signal.

2.0.0.3 Quantifying Coordination and Alignment

Coordination costs are a foundational concern in collaborative systems [28], [29]. However, evaluations of LLM multi agents often emphasize final outcomes (e.g., pass rates on benchmarks) over process efficiency. In software engineering, SWE-bench exposes the gap between realistic issue resolution and model capabilities [30], while SWE-agent [31] and OpenHands [15] reveal the importance of tool use and orchestration—but still provide limited real-time signals about team alignment. Complementary efforts argue for richer process traces and interactive environments to analyze coordination choices [16], [17], and work on LLM-as-judge warns that automated evaluators may introduce bias at the process level [32]–[34]. We contribute the Alignment Factor (AF) as a dynamic, real-time proxy for shared task understanding that both modulates work efficiency and directly informs agents’ communication decisions, thereby connecting conversational actions to tangible task outcomes.

3 C2C Framework↩︎

Communication to Completion is a time-stepped framework for deciding when, whom, and how to communicate in multi-agent collaboration. Figure 2 provides an overview of its mechanism using a software task. We organize the discussion as follows: §3.1 introduces the Sequential Action Framework that ensures deterministic execution, §3.2 describes the Alignment Factor mechanism for quantifying task understanding, §3.3 details how agents make communication decisions, §3.4 explains hierarchical task decomposition and tracking, and §3.5 presents the intention-based agent decision making that integrates these components.

3.1 Sequential Action Framework↩︎

Concurrent multi-agent execution often introduces temporal ambiguity. The Sequential Action Framework (SAF) constrains each agent to exactly one action per timestep, yielding deterministic state transitions and reproducible analysis.

3.1.0.1 Actions

SAF defines four actions capturing common collaborative behavior that capture common collaborative behavior: work (task execution with time cost), communicate (compose/send a message), reply (respond to an incoming message), and meeting (synchronous group discussion). Each action is temporally bounded and commits a single transition.

3.1.0.2 Synchronized Timesteps

All agents share a fixed temporal grid. At timestep \(t\), agent \(i\) selects and executes \(a_i^t\), then the system advances to \(t{+}1\). Formally, all actions at \(t\) finish before any action at \(t{+}1\) begins: \[\forall i,j:\; a_i^t \;\text{completes before}\; a_j^{t+1}\;\text{starts.}\]

3.1.0.3 Forward-only Delivery.

Messages sent at \(t\) are delivered at \(t{+}1\), preventing instantaneous feedback loops and enforcing causal consistency. A communication buffer stages pending deliveries and flushes at timestep boundaries.

3.2 Alignment Factor↩︎

The Alignment Factor (AF) measures an agent’s task-specific understanding and directly modulates effective productivity. Unlike static capability scores, AF evolves through communication. For agent \(i\) on task \(j\), \(AF_{i,j}\in[0.01,1.00]\) denotes comprehension quality. Values near \(0.01\) indicate limited understanding; values near \(1.00\) indicate mastery. We initialize \(AF_{i,j}{=}0.3\) to reflect the initial understanding.

When agent \(i\) receives reply about task \(j\) from agent \(k\), AF is updated based on the communication’s impact on task understanding. An LLM evaluates the message content, agent \(i\)’s current task context, and the relevance of the received information to determine the new alignment: \[AF_{i,j}^{\text{new}}=\min\!\bigl(1.0,\; AF_{i,j}^{\text{old}} + \Delta_{\text{eval}}\bigr),\] where \(\Delta_{\text{eval}} \in [0, 0.5]\) is computed by the LLM based on: (i) how well the message addresses the agent’s knowledge gaps, (ii) the relevance of the information to the specific task requirements, and (iii) the clarity and actionability of the guidance provided. Different message types naturally yield different impacts, help requests addressing critical blockers typically produce larger alignment gains.

If agent \(i\) spends \(h\) hours on task \(j\), the effective progress is \[\text{EffectiveProgress} \;=\; h \cdot AF_{i,j},\] A low AF impedes an agent’s progress, creating a natural incentive for it to seek clarification and coordination before investing substantial effort.

3.3 Agent Communication Decisions↩︎

Agents autonomously decide when, with whom, and how to communicate based on their current task state and situational awareness, without relying on fixed thresholds or rigid scheduling. For clarity, we decompose this decision process into four key components: initiating the communication, selecting the recipients, choosing the appropriate channel, and composing the message content.

3.3.0.1 Communication Initiation

Agents may initiate communication in response to various task situations: encountering technical difficulties (help_request), facing unclear requirements (need_clarification), identifying coordination needs with team members (meeting_invite), or reaching milestones that warrant status updates (progress_update). The decision to communicate emerges from the agent’s assessment of its current alignment and task complexity.

3.3.0.2 Recipient Selection

When initiating communication, agents consider multiple factors in selecting recipients: skill relevance to the issue, existing task dependencies, and historical interaction. The selection process reflects realistic collaboration patterns observed in human teams.

3.3.0.3 Channel Selection

Agents choose among three communication channels. Chat suits quick questions and simple clarifications with rapid response times. Email handles detailed explanations with longer response windows. Meetings address complex coordination needs requiring synchronous discussion among multiple participants.

3.3.0.4 Message Composition

Communication content is generated based on the agent’s current context, including task description, encountered difficulties, current alignment factor, and relevant technical details. Messages are composed to be informative and actionable, providing recipients with sufficient context to offer effective assistance.

3.4 Hierarchical Task Management↩︎

Complex tasks are decomposed by manager agents and tracked as a directed acyclic graph (DAG) of subtasks and dependencies.

The manager performs task decomposition in the beginning of the simulation, by analyzing the task requirements and proposing several subtasks according to team size and skills. Subtasks are connected through dependency edges. The manager tracks these dependencies and coordinates task assignments to ensure workers receive suitable subtasks. And the manager updates the task graph as workers complete subtasks, monitoring overall progress. Parent task progress is computed as a weighted average of subtask progress: \[\text{ParentProgress}=\frac{\sum_{i\in\text{subtasks}} w_i\,P_i}{\sum_{i\in\text{subtasks}} w_i},\] where \(w_i\) represents the effort estimate for subtask \(i\). The parent task is marked complete only when all required subtasks reach done status, ensuring accurate project tracking.

3.5 Intention-Based Agent Decision Making↩︎

Each agent makes decisions through an intention-based mechanism that evaluates the current situation and generates contextually appropriate actions.

3.5.0.1 Context Formation

At each timestep, the agent constructs a comprehensive context that includes: (i) currently assigned tasks and their completion status, (ii) alignment factors for each assignment indicating task understanding, (iii) recent communications including pending requests and received guidance, and (iv) team state including teammate skills and availability

3.5.0.2 Intention

Given this context, the agent uses an LLM to carefully reason about the current situation and generate an intention, a high-level decision about what action to take next. The generated intention is then translated into a concrete action within the SAF framework:

• Work: Continue task execution with effectiveness modulated by current AF.

• Help: Compose the message identifying difficulties and required expertise (HELP_REQUEST).

• Clarification: Request additional task details or requirement explanations (NEED_CLARIFICATION).

• Coordination: Propose synchronous discussion for complex collaboration needs (MEETING_INVITE).

• Reporting: Share subtasks progress with manager (PROGRESS_REPORT).

When generating communication intentions, the agent also determines appropriate recipients and selects suitable channels (chat, email, meeting) based on urgency and complexity.

3.5.0.3 Adaptive Behavior

This intention-based method enables agents to exhibit adaptive behavior without hard coded rules. The communication patterns emerge naturally from the agents’ contextual reasoning rather than threshold triggers.

By integrating these components, C2C enables agents to collaborate strategically and adaptively based on actual task needs.

4 Experiments↩︎

4.1 Experimental Setup↩︎

To evaluate the performance of C2C on realistic tasks, we implement experiments on software engineering workflows across three complexity tiers: Simple (8 hours, basic SWE operations), Medium (24 hours, API integration with authentication), and Complex (40 hours, full stack user authentication system); see Appendix 6 for details. Each task requires diverse skills including front-end development, back-end systems, database management, security implementation, and testing. Tasks are decomposed into several subtasks according to task complexity and team size with explicit skill requirements and dependencies, mirroring real-world software development scenarios.

4.1.0.1 Team Configurations

We systematically vary team sizes from 5 to 17 agents: 1M+4W (5 agents), 1M+8W (9 agents), and 1M+16W (17 agents), where M denotes manager and W denotes worker. Additionally, we evaluate multi-task scenarios with 1M+8W handling 2 concurrent tasks to assess workload distribution and task interference effects. Each agent is assigned 2–4 complementary skills from a skill pool, including skills like frontend, backend, and database.

4.1.0.2 Models and Implementation

All agents are powered by GPT-4o with temperature 0.7 for human like decision making. The Sequential Action Framework operates with 0.25 hour time steps over a maximum of 160 simulation steps (40 hours). Communication costs are calculated using realistic time models: email drafting (9 minutes base + content length), chat messages (3 minutes base), meetings (30 minutes minimum + preparation time + number of participants). We implement thread-depth limits (maximum 3 reply rounds) to prevent infinite communication loops.

4.1.0.3 Baselines

We compare C2C against two systematic baselines: (1) No Communication: agents work independently with task assignments but no communication, representing traditional parallel processing approaches; (2) Fixed Steps: agents communicate every 16 steps, simulating conventional project management practices. Both baselines use identical task decomposition and skill matching but lack C2C’s intelligent communication strategies.

4.1.0.4 Metrics

Following our evaluation framework, we report: task completion rate (percentage of tasks completed within time budget), average completion time (hours to finish successful tasks), communication cost (total time spent on communication activities), alignment score (average agent task alignment factor), and efficiency (ratio of productive work to total time investment). Each configuration is evaluated across all three task complexity tiers.

4.2 Main Results↩︎

As shown in 1, which details the results for a 1M+4W team, C2C consistently demonstrates superior performance in key efficiency metrics, although all methods achieved a 100% task completion rate. The primary advantage of C2C is most evident in task completion time. On Complex tasks, C2C finishes in 24.75 hours, significantly faster than both No Communication (33.5 hours) and Fixed Steps (36.25 hours). This trend holds for Medium tasks, where C2C is also the fastest.

This time saving translates to higher efficiency. For Complex tasks, C2C’s efficiency score is 1.62, substantially outperforming Fixed Steps (1.10) and No Communication (1.19). Similarly, C2C achieves the highest alignment score on Medium (0.68) and Complex (0.55) tasks, confirming that its communication strategy effectively enhances agent understanding. While communication costs are comparable across strategies, C2C utilizes its communication budget more effectively to achieve better outcomes.

Figure 3 illustrates the dynamics of the alignment factor mechanism. Alignment scores start low as agents have minimal task understanding, then improve through strategic communications. Help requests and meetings yield the strongest alignment improvements, while progress updates are modest. The communication heatmap reveals the manager centric communication pattern.

4.3 Scalability and Multi-Task Analysis↩︎

Table 2 demonstrates C2C’s effectiveness across team sizes and workload scenarios. The framework exhibits sub-linear scaling in communication cost: while team size increases to 3.4 times (5 to 17 agents), communication cost only increases 86%. This favorable scaling behavior stems from C2C’s intelligent communication routing and the alignment factor mechanism, which prioritizes high value interactions over communications.

Speedup analysis reveals that performance steadily improves as the number of agents increases, with the largest configuration tested (1M+16W) achieving the highest speedup of \(\mathbf{1.95 \times}\). This highlights the effective task parallelization and specialized skill utilization in larger teams. However, the incremental gain in speedup lessens as the team grows, suggesting that simply adding more workers yields progressively smaller advantages.

The multi-task evaluation with 1M+8W reveals C2C’s ability to handle concurrent workloads effectively. When processing 2 tasks simultaneously, completion time increases from 11.25h to 21h (an 87% increase), significantly better than naive linear scaling. Analysis of communication patterns shows that C2C naturally evolves hub-and-spoke topologies with managers as primary coordinators, avoiding the quadratic communication complexity that plagues peer-to-peer approaches. In multi-task scenarios, agents exhibit sophisticated context switching behavior, maintaining separate alignment factors per task and prioritizing communications based on overall workflow optimization.

4.4 Message Type and Decomposition↩︎

4.4.0.1 Message Type Analysis

Table 3 details the impact of different message types on collaboration. MEETING_INVITE messages provide the highest alignment gains (+0.27), followed by HELP_REQUEST (+0.15), directly correlating with task success by resolving critical blockers. In contrast, PROGRESS_UPDATE messages maintain synchronization with a neutral (0) impact on alignment in this context. This analysis validates C2C’s strategy of prioritizing communications that address specific knowledge gaps.

4.4.0.2 Task Decomposition Quality

Effective collaboration begins with high quality task decomposition. We compare C2C’s adaptive decomposition strategy against manual, naive LLM, and hierarchical methods in Table 4. C2C’s approach, which considers agent skills and workload during decomposition, achieves superior subtask clarity and higher AF. This leads to 14% increase in worker utilization rate compared to naive LLM decomposition, demonstrating that context aware planning is critical for multi-agent efficiency.

4.5 Communication Pattern Analysis↩︎

Figure 4 shows how communication varies with task complexity and team size (1M+4W, 1M+8W, 1M+16W). (a) Completion time increases with task complexity; adding workers shortens time but with diminishing returns. (b) Message intensity (messages/hour) rises with complexity and team size.

4.5.0.1 Message Types

In (c), simple tasks are dominated by help and clarification (about 66% and 34%, respectively). At medium complexity, help remains the majority (\(\approx\)​83%), while meetings appear (\(\approx\)​8%). For complex tasks, the mix diversifies: help \(\approx\)​70%, and meetings \(\approx\)​13% become dominant. C2C allocates most messages to high value help requests and escalates to meetings only when the expected coordination benefit exceeds the cost.

4.5.0.2 Channels and Latency

Panel (d) indicates a shift from chat toward email as complexity grows, with a modest rise in meetings. The per channel response times for chat and email stay roughly flat across complexity levels (reduce slightly), whereas meeting latency increases for complex tasks.

These patterns align with the logic of C2C: as tasks become more complex, agents continue to seek help most of the time but increasingly use meetings, while clarification needs drop. The engine selects higher yield (though costlier) channels when needed, while larger teams reduce completion time without eliminating the upward pressure on communication volume with complexity.

5 Conclusion↩︎

In this paper, we present Communication to Completion, a multi-agent framework that treats communication as an optimizable resource and quantifies task understanding through the Alignment Factor within a Sequential Action Framework. By treating alignment as a quantifiable and dynamic variable, C2C enables agents to autonomously determine when and how to communicate.

Our experiments demonstrate that this alignment-driven approach adapts naturally to task complexity, with agents communicating more frequently and through richer channels when facing difficult problems while minimizing overhead on simpler tasks. Across diverse task complexities and team sizes, C2C consistently achieves high task completion while reducing completion time and improving efficiency relative to baselines. The framework produces interpretable patterns, including alignment factor trajectories, communication type distributions, and collaboration network structures, which bridge agent behavior and coordination theory to provide design insights for multi-agent systems.

Limitations↩︎

While the C2C framework demonstrates significant performance gains, we acknowledge several limitations that offer avenues for future research. The findings are derived from a controlled simulation, and the framework’s performance in unpredictable, real-world environments is yet to be validated. The experiments were confined to the software engineering domain, so the C2C’s effectiveness may not generalize to other collaborative fields. Furthermore, the Alignment Factor (AF) relies on an LLM’s evaluation, which introduces potential subjectivity and bias.

6 Tasks Details↩︎

This part provide example task prompts used in our simulations. Each prompt specifies the description, time budget, and required skills.

7 Implementation Details↩︎

We provide prompts used in our experiments in this section.

References↩︎