SkillFlow: Efficient Skill and Code Transfer Through Communication in Adapting AI Agents

TL;DR

SkillFlow introduces a decentralized, on-demand skill-transfer framework for AI agents, enabling agents to acquire and integrate new capabilities from peer providers through a skill selection, communication, and integration pipeline. The approach is validated with both a simulation-based cost model and an application-based calendar scheduling benchmark, showing favorable cost and time reductions, particularly when communication costs dominate, and with initial buy costs amortized over iterations. The work draws parallels to lateral gene transfer to frame adaptation and discusses trade-offs between centralized marketplaces and decentralized systems, along with security and future research directions. The results suggest SkillFlow can meaningfully accelerate task completion and reduce cumulative costs in dynamic, multi-agent settings, with practical impact for scalable autonomous agent ecosystems.

Abstract

AI agents are autonomous systems that can execute specific tasks based on predefined programming. Here, we present SkillFlow, a modular, technology-agnostic framework that allows agents to expand their functionality in an ad-hoc fashion by acquiring new skills from their environment or other agents. We present a theoretical model that examines under which conditions this framework would be beneficial, and we then explore SkillFlow's ability to accelerate task completion and lead to lower cumulative costs in a real-world application, namely scheduling agents for calendar events. We demonstrate that within a few iterations, SkillFlow leads to considerable (24.8%, p-value = $6.4\times10^{-3}$) gains in time and cost, especially when the communication cost is high. Finally, we draw analogies from well-studied biological systems and compare this framework to that of lateral gene transfer, a significant process of adaptation and evolution in novel environments.

Figures (7)

• Figure 1: Overview of SkillFlow. A. Three potential pathways for skill sharing are (1) direct skill sharing between agents, (2) acquiring skills from a centralized database or marketplace, and (3) enhancing an existing skill through self-improvement. In this work, SkillFlow focuses on the first pathway. B. The architecture of the SkillFlow module. When the user agent (top) receives user instructions, it attempts to identify and select relevant skills. If the agent lacks the required skill, it queries its local skill register to locate a source agent (bottom) that possesses the needed skills. Note that the skill register is a decentralized, peer-to-peer storage per user, which differs from the centralized skill database. The communication module facilitates negotiation and transaction of the skill between the two agents. Once acquired, the user agent integrates the new skill and executes the task with its enhanced capabilities. Additionally, agents can log new skills in the skill register at any time, expanding the skill-sharing network over time.
• Figure 2: Simulation-based benchmark results for SkillFlow. A-C, Triangle heatmaps showing the average cost-per-task difference between the SkillFlow$_\$$ and baseline across different iterations. Each point represents a simulation parameter set consisting of three cost components (Buying, Execution, and Communication) that always sum to 20. The closer a point to a corner, the higher that specific cost parameter. Green areas indicate scenarios where SkillFlow$_\$$ is more cost-efficient than the baseline. Black dots represent simulations using the parameter set (Buying = 14, Execution = 2, Communication = 4). Each parameter set was repeated with 10 different random seeds, and the average result was taken. D, Trajectory plot showing the average cost per task over iterations for different scenarios using the parameter set indicated by black dots in the triangle heatmaps. The error band represents the 95% confidence interval computed over 10 runs. E, Trajectory plot illustrating the changes in the average cost-per-task difference between SkillFlow$_\$$ and baseline with regard to communication-to-execution cost ratios while keeping the buying cost fixed to 4. The error bars indicate the 95% confidence interval computed over 10 runs.
• Figure 3: Design of the application-based benchmark for SkillFlow. A, Skills are initially distributed across different agents (left panel). The user can only provide instructions to Agent 1, and this agent can interact and acquire skills from Agents 2 and 3. After a certain number of iterations, the user can now accomplish tasks much more efficiently because Agent 1 no longer requires dependencies of other agents (right panel). B, Workflow of a calendar scheduling agent that uses SkillFlow to enhance its capabilities. The agent initially lacks traffic or weather forecasting functionality. With SkillFlow, after acquiring the forecasting capabilities, the agent can now utilize these skills to handle more complex user instructions, such as “Go for a coffee with Bob where the traffic is light.” After using these skills to enrich its context, the agent communicates with another user or agent to finalize the scheduling.
• Figure 4: Demonstration of calendar agent with SkillFlow. Dialogue between the user and agents to schedule a coffee meetup with Bob. The user (Mary) requests that the calendar agent obtain information about a local coffee shop. The user's assistant uses SkillFlow to acquire the necessary skills from another agent (Helen's Assistant). This skill is then used to enhance the context for the user's agent. Next, the user's agent communicates with Bob's agent until both agree on the meetup time. Finally, the user's agent updates the calendar and reports back to the user, concluding the task.
• Figure 5: Application-based benchmark results for SkillFlow. Time trajectory of the average time per task for all completed tasks up to each iteration. The bar plot represents the p-values where SkillFlow significantly differed from the baseline at each iteration. P-values were computed using a two-tailed t-test and adjusted for false discovery rate. The line plot at the bottom shows the percentage of new skills learned up to each iteration. The error band represents the 95% confidence interval computed over 20 runs.