Communication to Completion: Modeling Collaborative Workflows with Intelligent Multi-Agent Communication

1 Introduction

• Teamwork requires weighing information-sharing value against meeting time, asynchronous delays, and interruptions that fragment attention.The introduction frames when, with whom, and how to communicate as a central challenge.Source: §1 · p. 1
• C2C offers a controlled simulation environment because real-team collaboration is difficult to study when variables cannot be controlled or scenarios replayed.The introduction positions multi-agent LLM systems as a way to investigate workplace collaboration patterns.Source: §1 · pp. 1–2
• C2C models communication costs explicitly, while AF measures shared task understanding and guides whether communication is worth its cost.AF is presented as a dynamic metric inspired by Shared Mental Models.Source: §1 · p. 2
• Cost-aware communication strategies outperform unconstrained interaction in efficiency on complex software engineering tasks.The introduction also reports that constraints induce more structured communication topologies.Source: §1 · p. 2

2.1 Multi-Agent Collaboration Frameworks

• Task-oriented frameworks solve engineering tasks but commonly treat interaction as low-friction messaging, while behavioral simulations model social dynamics with less focus on goal-directed engineering.The paper positions its work between these directions by targeting task-oriented collaboration under explicit coordination constraints.Source: §2.1 · p. 2

2.2 Communication Costs and Constraints

• Prior LLM-agent frameworks rarely model asynchronous delay or context-switch overhead, despite evidence that coordination costs can dominate delivery time in distributed software teams.The paper distinguishes its discrete-event simulation engine from systems that treat communication as a byproduct of reasoning.Source: §2.2 · p. 2

2.3 Shared Mental Models and Alignment

• Outcome-correctness measures and interaction-trajectory analyses do not provide a dynamic measure of consensus during collaboration; AF is introduced as a computational proxy for shared mental models.AF is used to quantify how communication updates shared context and how evolving alignment relates to execution efficiency.Source: §2.3 · pp. 2–3

3.1 Discrete Event Simulation Engine

• The discrete-event engine gives each agent one action per timestep, advances agents on a shared grid, and delivers messages at the next timestep for deterministic, causal execution.Actions include work, communication, replies, and meetings.Source: §3.1 · p. 3

3.2 Alignment Factor

• AF tracks task-specific understanding, starts at 0.3, updates from communication quality, and scales effective progress; low AF impedes work and motivates clarification despite its cost.The simulation evaluates communication quality using knowledge-gap resolution, task relevance, and actionability.Source: §3.2 · pp. 3–4

3.3 Agent Communication Decisions

• Agents initiate help or meeting requests for task difficulties or dependencies, select recipients by skill relevance and history, and choose channels using task fit, communication costs, and channel preference.Channel utility weighs bandwidth suitability against latency and synchronization overhead, alongside agent-specific preferences.Source: §3.3 · p. 4

3.4 Hierarchical Task Management

• Managers decompose complex tasks into a dependency DAG, assign suitable subtasks, update it as work finishes, and compute effort-weighted parent progress; completion requires every required subtask to be done.Subtask weights represent effort estimates.Source: §3.4 · p. 4

3.5 Intention-Based Agent Decision Making

• LLM intentions use assigned tasks, AF, recent communications, and team state to guide work, help, coordination, or reporting actions.The agent constructs this context at each timestep before generating an intention.Source: §3.5 · pp. 4–5
• C2C achieves the highest Efficiency on Medium and Complex tasks, while Free Comm is faster in raw completion but incurs high communication costs.Efficiency is task progress normalized by total time investment; reported values are means ± standard deviations over five runs.Source: Table 1 · p. 5
• Agents select recipients and chat, email, or meeting channels according to urgency and complexity, with communication patterns emerging from contextual reasoning rather than hard-coded thresholds.The intention-based approach is described as enabling strategic, adaptive collaboration based on task needs.Source: §3.5 · p. 5

4.1 Experimental Setup

• The evaluation covers 15 software-engineering workflows across three complexity tiers, with teams of 5, 9, or 17 agents and metrics spanning completion, communication cost, alignment, and efficiency.Baselines are No Communication, Fixed Steps, and unconstrained Free Communication; communication activities have modeled time costs.Source: §4.1 · pp. 5–6

4.2 Main Results

• C2C records Efficiency 1.85 on Medium and 1.62 on Complex tasks; on Complex tasks, communication cost is 7.02 hours versus Free Communication’s 14.55 hours.Figure 3 reports alignment rising across rounds, with high-intent communications yielding the largest gains and the manager acting as a central hub.Source: §4.2 · p. 6, Figure 3 · p. 6

4.3 Generalization across Models

• C2C improves Efficiency across GPT-5.2, Claude Sonnet 4.5, and Gemini 2.5 Pro; for Claude Sonnet 4.5, Efficiency rises from Free Communication’s 1.96 to 2.40, a 22% relative improvement.The table caption describes the comparison as applying to Medium-complexity tasks.Source: §4.3 · pp. 6–7, Table 2 · p. 7

4.4 Scalability and Multi-Task Analysis

• As teams grow, incremental speedup gains lessen, with reported speedup ranging from 1.54× to 1.95×; handling two concurrent tasks raises completion time from 11.25h to 21h (87%).Speedup is relative to the No Communication baseline; the reported two-task completion-time increase is described as better than naive linear scaling.Source: §4.4 · p. 7, Table 3 · p. 7

4.5 Message Type and Decomposition

• MEETING_INVITE messages yield the largest reported alignment gain (+0.27), followed by HELP_REQUEST (+0.15); PROGRESS_UPDATE has negligible immediate impact but maintains long-term context.Source: §4.5 · p. 7, Table 4 · p. 7
• C2C’s adaptive task decomposition achieves subtask clarity 0.95, above naive LLM prompting at 0.72 and near human manual quality at 1.00.Source: §4.5 · p. 7, Table 5 · p. 7

4.6 Communication Pattern Analysis

• Across team configurations, completion time rises with task complexity while larger worker pools shorten it with diminishing returns; message intensity rises with complexity and team size.The communication mix shifts with complexity: help remains the majority, while meetings become more prominent; C2C uses meetings when expected coordination benefit exceeds cost.Source: §4.6 · pp. 7–8, Figure 4 · p. 8
• As task complexity grows, communication shifts from chat toward email, with a modest increase in meetings; meeting latency also increases for complex tasks.Chat and email response times stay roughly flat across complexity levels, reducing slightly.Source: §4.6 · p. 8, Figure 4 · p. 8

• C2C formalizes task understanding with the Alignment Factor and models communication as a scarce resource with explicit latency to quantify coordination-cost and execution-efficiency trade-offs.Source: §5 · p. 8
• Experiments report strategic behaviors such as batching questions and escalating to meetings, with patterns consistent across different backbone models.Source: §5 · pp. 8–9
• The Alignment Factor does not guarantee functional correctness, deterministic costs simplify stochastic interruptions, and experiments are limited to software engineering.Generalization to other collaborative domains remains future work.Source: §5 · p. 9

• Example task prompts specify a task description, time budget, and required skills.Source: A Tasks Details · p. 11

• The examples include an 8-hour task to fix five independent bugs across modules, with no cross-dependencies.Source: B Implementation Details · p. 11
• A 24-hour example integrates an external API with authentication, token management, and error or latency handling.Source: B Implementation Details · p. 11
• The appendix also shows prompts for manager task decomposition, agent intention generation, and Alignment Factor updates.Source: Figure 5 · p. 12, Figure 6 · p. 12, Figure 7 · p. 13