PODCAST — Claude Dynamic Workflows: Scaling Complex Work Through Orchestration

PODCAST — Claude Dynamic Workflows: Scaling Complex Work Through Orchestration

Prefer reading instead? The full article is available here [https://open.substack.com/pub/aipractitioner/p/claude-dynamic-workflows-scaling?r=49ttp&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true]. The podcast is also available on Spotify [https://open.spotify.com/show/6MROBKvrjx0Mey8tHud5LX] and Apple Podcasts [https://podcasts.apple.com/us/podcast/the-ai-practitioner-podcast/id1830285899]. Subscribe to keep up with the latest drops. LLMs excel at individual tasks. But most valuable work isn’t a single task, it’s a coordinated process. Writing a research report, reviewing a large codebase, or evaluating conflicting evidence requires multiple stages of analysis, validation, and synthesis without losing rigor along the way. In this episode, we explore Claude Dynamic Workflows, the multi-agent orchestration framework Anthropic released on May 28, 2026. Rather than forcing everything through a single conversation, Claude generates a JavaScript workflow that defines the phases of work, determines what can run in parallel, and routes information between stages. You’ll learn: * Where dynamic workflows fit on the coordination ladder: why single agents often struggle with laziness, self-preferential reasoning, and goal drift, and when a workflow outperforms a skill, a subagent, or a full agent team. * What happens under the hood: how Claude transforms a prompt into an inspectable execution script, how isolated agents fan out to investigate different aspects of a problem, and how six recurring workflow patterns power most real-world use cases. * What a real workflow looks like in practice: through the example of a deep-research workflow on developer productivity, we'll see how Claude spawned 27 agents, spent roughly $5, and used an independent verification phase to explain a genuine conflict in the literature instead of averaging it away. If you’d rather read than listen, the full article (with diagrams, code examples, and implementation details) is available on Substack: 👉 Enjoyed this episode? Subscribe to The AI Practitioner to get future articles and podcasts delivered straight to your inbox: aipractitioner.substack.com [https://aipractitioner.substack.com/] This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit aipractitioner.substack.com [https://aipractitioner.substack.com?utm_medium=podcast&utm_campaign=CTA_1]

PODCAST — SkillOpt Explained: From Prompt Engineering to Skill Training

Prefer reading instead? The full article is available here. The podcast is also available on Spotify [https://open.spotify.com/show/6MROBKvrjx0Mey8tHud5LX] and Apple Podcasts [https://podcasts.apple.com/us/podcast/the-ai-practitioner-podcast/id1830285899]. Subscribe to keep up with the latest drops. Skills have become a core building block for AI agents, packaging expertise and workflows into reusable instructions. But despite their growing importance, most teams still create them through trial and error: write a skill, test it, revise it, and hope it performs better. In this episode, we explore SkillOpt, a framework that treats skills not as static prompts but as trainable artifacts. Instead of manually rewriting instructions, SkillOpt uses evaluation, feedback, and optimization loops to systematically improve a skill over time. You’ll learn: * How SkillOpt reframes prompt engineering as skill training: why skills can be treated as trainable external state and how optimization happens without changing model weights. * What the experiments reveal about skill optimization: where the largest gains appear, why skill transfer works across models, and the practical limits of optimizing instructions instead of model parameters. * How to apply SkillOpt yourself: reproducing the workflow on SearchQA, training skills on your own datasets, and evaluating whether optimization delivers measurable improvements over baseline prompting. By the end, you’ll understand when skill optimization is worth the additional compute, where it outperforms traditional prompt engineering, and why training instructions may become as important as training models in future agent systems. If you’d rather read than listen, the full article (with diagrams, code examples, and implementation details) is available on Substack: 👉 Enjoyed this episode? Subscribe to The AI Practitioner to get future articles and podcasts delivered straight to your inbox. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit aipractitioner.substack.com [https://aipractitioner.substack.com?utm_medium=podcast&utm_campaign=CTA_1]

PODCAST — Score-Based Causal Discovery with LLMs

Prefer reading instead? The full article is available here [https://open.substack.com/pub/aipractitioner/p/score-based-causal-discovery-with?r=49ttp&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true]. The podcast is also available on Spotify [https://open.spotify.com/show/6MROBKvrjx0Mey8tHud5LX] and Apple Podcasts [https://podcasts.apple.com/us/podcast/the-ai-practitioner-podcast/id1830285899]. Subscribe to keep up with the latest drops. Constraint-based methods discover causal graphs by testing independences. Score-based methods take a different route. They treat causal discovery as a model selection problem, scoring candidate graphs and searching for the one that best balances fit and complexity. In this episode, we explore how score-based algorithms learn causal structure, why they hit the same identifiability ceiling as constraint-based methods, and how LLMs can be plugged into the search itself rather than just bolted on at the end. You’ll learn: * How score-based methods differ from constraint-based ones: why framing causal discovery as model selection changes both the search procedure and the kinds of errors the algorithm makes. * Where LLMs can intervene in score-based pipelines: the five integration points, from hard constraints to iterative agentic loops, and which ones are recoverable when the LLM is wrong. * How to pick the right algorithm and LLM integration strategy: comparing priors, post-hoc orientation, and score augmentation on the Adult Census Income dataset, and what each one is worth in practice. By the end, you’ll have a clear view of where score-based methods sit relative to constraint-based ones, and a practical map of how to combine statistical search with LLM-derived priors without letting the LLM override the data. If you’d rather read than listen, the full article (with diagrams, code examples, and implementation details) is available on Substack: 👉 Enjoyed this episode? Subscribe to The AI Practitioner to get future articles and podcasts delivered straight to your inbox. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit aipractitioner.substack.com [https://aipractitioner.substack.com?utm_medium=podcast&utm_campaign=CTA_1]

PODCAST — LLMs in Causal Discovery: A Deep Dive into Constraint-Based Algorithms

Prefer reading instead? The full article is available here. The podcast is also available on Spotify [https://open.spotify.com/show/6MROBKvrjx0Mey8tHud5LX] and Apple Podcasts [https://podcasts.apple.com/us/podcast/the-ai-practitioner-podcast/id1830285899]. Subscribe to keep up with the latest drops. Most ML models answer one question: what is likely to happen? The harder question is what will change if you intervene. That gap is where causal reasoning begins. In this episode, we explore how constraint-based algorithms learn causal structure directly from data, and how LLMs can step in to resolve what statistics alone cannot. You’ll learn: * How PC, FCI, and RFCI discover causal graphs using conditional independence tests, and what assumptions each one makes. * How to encode domain knowledge as hard constraints, so the algorithm stops producing edges that are statistically plausible but practically nonsensical. * How LLMs can review and refine the output graph, resolving ambiguous orientations with domain reasoning when the data runs out of signal. By the end, you’ll have a clear picture of a three-layer pipeline that combines statistical discovery, expert constraints, and LLM review into a coherent approach to causal graph learning. If you’d rather read than listen, the full article (with diagrams, code examples, and implementation details) is available on Substack: 👉 Enjoyed this episode? Subscribe to The AI Practitioner to get future articles and podcasts delivered straight to your inbox. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit aipractitioner.substack.com [https://aipractitioner.substack.com?utm_medium=podcast&utm_campaign=CTA_1]

PODCAST — Building Claude Skills: A New Paradigm for Interacting with LLMs

Prefer reading instead? The full article is available here [https://aipractitioner.substack.com/p/long-term-memory-unlocking-smarter-38d]. The podcast is also available on Spotify [https://open.spotify.com/show/6MROBKvrjx0Mey8tHud5LX] and Apple Podcasts [https://podcasts.apple.com/us/podcast/the-ai-practitioner-podcast/id1830285899]. Subscribe to keep up with the latest drops. Large language models are powerful, but relying on prompts alone quickly becomes fragile and difficult to scale. As teams try to operationalize LLMs in real workflows, traditional documentation and ad-hoc prompting start to break down. In this episode, we explore a new paradigm introduced with Claude Skills: packaging workflows, instructions, and resources into reusable capabilities that LLMs can execute. You’ll learn: * Why traditional documentation is poorly suited for LLMs and why workflow-first instructions are more effective. * How Claude Skills structure tasks using a concise SKILL.md file that points to supporting files and scripts loaded on demand. * How teams can design and deploy skills to turn LLMs into reliable task executors rather than prompt-driven tools. By the end, you’ll understand how skills move us from prompt engineering to designing AI-native workflows. If you’d rather read than listen, the full article (with diagrams, code examples, and implementation details) is available on Substack: 👉 Enjoyed this episode? Subscribe to The AI Practitioner to get future articles and podcasts delivered straight to your inbox. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit aipractitioner.substack.com [https://aipractitioner.substack.com?utm_medium=podcast&utm_campaign=CTA_1]