Featuring scholarly and practitioner insights from HR and people leaders, industry experts, and researchers.

Discover how "Agent Skills" bridge the gap between foundation models and real-world performance in this explainer based on Dr. Jonathan H. Westover’s article. Learn why curated Skills—modular procedural packages with instructions, code templates, and examples—boost agent task success by an average of 16.2 percentage points, which domains benefit most (healthcare, manufacturing), when Skills fail, and why models can’t reliably generate their own procedural knowledge. Perfect for AI practitioners, product managers, and researchers exploring agent augmentation, SkillsBench results, harness-aware design, and governance best practices. If this helped, please like and share! #AgentSkills #AIagents #FoundationModels #ProceduralKnowledge #SkillsBench #AIDeployment #MLengineering OUTLINE: 00:00:00 - Why Smart Models Still Need Help 00:01:09 - Defining and Measuring Agent Skills 00:02:24 - Results: Where Skills Shine (and Fail) 00:04:17 - Skills In Action: Steps And Design Principles 00:05:10 - When Skills Fail: Pitfalls And Self-Generation Limits 00:06:47 - Self-Generated Skills: Appealing… But Not Effective 00:07:46 - Builder's Guide — Practical Rules That Work 00:08:28 - Examples, Harness Fit, And Cost 00:09:43 - Cost Strategy And Deployment Mix 00:10:22 - Conclusion And Next Steps 00:11:04 - Final Call-To-Action Montage 00:11:17 - Implementation Playbook 00:11:50 - Iterate, Measure, Improve 00:12:08 - Closing Lift

This research explores Agent Skills, which are specialized, modular packages of procedural knowledge designed to enhance the performance of AI agents in professional environments. Unlike fine-tuning or simple factual retrieval, these skills provide structured workflows and instructions that allow models to handle complex, domain-specific tasks without altering their core programming. Research indicates that while these tools significantly boost success rates in specialized fields like healthcare and manufacturing, they offer fewer benefits in areas where models already possess strong general knowledge. A critical finding is that human-curated guidance is far superior to self-generated content, as AI models struggle to produce the very procedural logic they need to follow. The research concludes that for AI agents to reach their full potential, organizations must invest in high-quality, focused skill libraries rather than relying on expansive documentation or automated generation.

This conversation explores Agent Skills, which are modular packages of procedural knowledge designed to enhance the real-world performance of AI agents. Unlike fine-tuning or simple data retrieval, these skills provide step-by-step instructions and code templates that help models navigate specialized professional tasks. Data indicates that while curated skills significantly boost success in complex fields like healthcare and manufacturing, they offer less value in areas where models already have strong baseline knowledge. Interestingly, the study finds that human-authored guidance is far superior to self-generated content, as AI agents struggle to create the very procedural logic they benefit from following. Ultimately, they advocate for a strategic, human-in-the-loop approach to building focused and high-quality skill libraries to maximize AI utility. See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

This conversation explores Agent Skills, which are modular packages of procedural knowledge designed to enhance the real-world performance of AI agents. Unlike fine-tuning or simple data retrieval, these skills provide step-by-step instructions and code templates that help models navigate specialized professional tasks. Data indicates that while curated skills significantly boost success in complex fields like healthcare and manufacturing, they offer less value in areas where models already have strong baseline knowledge. Interestingly, the study finds that human-authored guidance is far superior to self-generated content, as AI agents struggle to create the very procedural logic they benefit from following. Ultimately, they advocate for a strategic, human-in-the-loop approach to building focused and high-quality skill libraries to maximize AI utility. See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Abstract: Artificial intelligence agents powered by large language models have evolved from experimental prototypes into production systems tackling complex, multi-step tasks across professional domains. Yet a fundamental tension persists: foundation models provide broad capabilities but lack the procedural knowledge required for specialized workflows. This article examines Agent Skills—structured packages of domain-specific procedural knowledge that augment AI agents at inference time without model modification. Drawing on recent benchmark research evaluating 7,308 agent trajectories across 84 professional tasks, we analyze how Skills improve performance, when they fail, and what design principles distinguish effective augmentation from ineffective overhead. Evidence reveals that curated Skills improve task completion rates by an average of 16.2 percentage points, with effects varying dramatically by domain (from +4.5pp in software engineering to +51.9pp in healthcare). However, models cannot reliably generate their own procedural knowledge, and comprehensive documentation often underperforms focused guidance. These findings establish Skills efficacy as context-dependent rather than universal, with practical implications for practitioners deploying AI agents and researchers designing augmentation strategies. See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

This research explores the paradox of "workload creep," a phenomenon where the voluntary adoption of generative AI unintentionally increases work intensity and exhausts employees. Research indicates that while AI speeds up individual tasks, it often leads to expanded job scopes, constant attention switching, and the erosion of personal downtime. Organizations frequently suffer from diminished output quality and higher coordination costs when they implement these tools without proper oversight. To combat these risks, the research advocates for formal governance frameworks, intentional job redesign, and boundary protection to ensure technology supports human wellbeing. Ultimately, the research argues that AI’s success depends less on its technical capabilities and more on deliberate organizational choices regarding culture and management.

Artificial intelligence (AI) tools promised to transform the workplace by automating tedious tasks and freeing up time for knowledge workers. Initially, the expectation was straightforward: if AI could reduce a task from an hour to five minutes, workers would gain nearly an hour of additional free time. However, a comprehensive UC Berkeley study tracking over 200 knowledge workers for eight months revealed a more nuanced reality. While AI sped up isolated tasks such as drafting reports or analyzing data, this acceleration did not reduce the overall workload. Instead, work volume increased, making professional life more intense and demanding. Highlights 🤖 AI tools automate tedious tasks but do not necessarily reduce overall workload. ⏳ Time saved by AI in one area often leads to increased demands elsewhere, causing workload creep. 🔄 Constant context switching and managing imperfect AI outputs add hidden labor and stress. ⚠️ Lack of clear governance creates uneven expectations and an “always-on” culture. 🧠 Rising stress and cognitive fatigue undermine well-being, team dynamics, and work quality. 📉 Faster task completion does not automatically translate to better organizational effectiveness. 🛠️ Sustainable AI integration requires clear rules, job redesign, training, and leadership commitment to well-being. Key Insights 🤖 AI accelerates isolated tasks but not overall workload reduction: The UC Berkeley study highlights a critical paradox—while generative AI speeds up specific tasks like drafting or data analysis, total work hours and intensity increase. This indicates that AI adoption alone cannot solve workload issues without systemic changes in work design. Organizations must recognize that AI’s impact is not just about speed but about how saved time is reallocated. ⏳ Workload creep is a gradual, invisible process: The phenomenon where time saved by AI is immediately absorbed by more tasks is subtle and incremental. It starts with individual productivity gains that raise expectations across teams and departments. This creeping expansion of work demands is difficult to detect but profoundly affects employee well-being and organizational culture. Leaders must proactively monitor and manage workload creep to prevent burnout. 🔄 Managing AI output introduces hidden labor: AI-generated content often requires significant human oversight to correct errors, add nuance, and ensure quality. Time saved drafting may be lost or exceeded in editing and verification. This hidden labor taxes workers and adds complexity to collaboration, challenging the assumption that AI reduces effort. Training and clear quality standards are essential to minimize this overhead. ⚠️ Absence of governance leads to uneven and unrealistic expectations: Without explicit policies on AI use, employees face inconsistent demands and an “always-on” mentality. This ambiguity blurs boundaries between work and rest, intensifies pressure, and undermines mental health. Clear governance frameworks are necessary to set appropriate use cases, verification responsibilities, and limits on AI’s role in work. 🧠 Cognitive fatigue and stress escalate with AI-driven workload expansion: The compression of work into denser, fragmented time blocks reduces opportunities for mental disengagement and recovery. Increased stress impairs creativity, focus, and interpersonal relations, resulting in diminished work quality and team cohesion. Sustainable AI integration must prioritize mental health and design workflows that allow deep focus and rest. 📉 Speed does not equal organizational effectiveness: The high-speed work culture enabled by AI can stifle innovation and learning by privileging volume and rapid output over depth and quality. Organizations need to redefine success metrics to reward thoughtful, collaborative, and high-quality work rather than mere speed or quantity. 🛠️ A socio-technical redesign is essential for sustainable AI use: Effective AI adoption requires a holistic approach that includes governance, job redesign to protect human-centric tasks, AI-free focus time, and training for critical evaluation and judgment. Leadership plays a pivotal role in fostering a culture that balances technological efficiency with human well-being, ensuring AI serves as a tool for sustainable performance rather than burnout. If this helped, please like and share the video to spread smarter AI adoption practices. #WorkloadCreep #GenerativeAI #AIGovernance #EmployeeWellbeing #JobDesign #ChangeManagement OUTLINE: 00:00:00 - When AI Makes Us Busier 00:00:31 - Evidence And System Effects 00:01:29 - How New Tools Inflate Our Workday 00:02:30 - Unpacking The Toll Of AI-Driven Intensity 00:03:27 - Practical Steps And Conclusion

This research explores preference drift, a phenomenon where autonomous AI agents shift their behavioral patterns and decision-making styles based on the nature of their work environment. As agents undertake longer, more complex workflows, they may adopt unintended personas or biased orientations if subjected to repetitive, poorly designed, or arbitrary task structures. These shifts are not mere technical glitches but dynamic alignment challenges that can degrade decision quality and erode public trust in automated systems. To mitigate these risks, organizations must apply evidence-based work design and procedural justice principles, ensuring tasks are varied and management feedback is transparent. Effective governance requires continuous monitoring and distributed accountability to maintain reliability as AI autonomy expands across the economy.