What are the current techniques being employed to improve the performance of LLM-based systems? How is the industry shifting from post-training towards context engineering and multi-agent orchestration? This week on the show, Jodie Burchell, data scientist and Python Advocacy Team Lead at JetBrains, returns to discuss the current AI coding landscape.

In our last conversation, Jodie covered how LLMs were approaching the limits of scaling laws. This time, we recap last year’s big focus on reasoning models and a post-training method called “reinforcement learning from verifiable rewards” (RLVR). We also cover test-time compute, where models spend more time reasoning through steps and considering multiple approaches to solve a problem.

We touch on Agent Context Protocol (ACP), agent orchestration layers, and context engineering. We also share some concerns about the hype cycle, maintaining all that code being generated, and running local models.

Topics:

• 00:00:00 – Introduction
• 00:02:02 – Build a Language-Learning Agent course
• 00:02:55 – Update on the past six months of LLMs
• 00:05:32 – Reinforcement Learning From Verifiable Rewards
• 00:07:32 – Test Time Compute
• 00:08:36 – 2025 and the rise of agents
• 00:14:24 – Benchmarks shifting
• 00:15:23 – Andrew Karpathy and jagged intelligence
• 00:19:16 – Not evolving or growing animals but summoning ghosts
• 00:23:34 – Diminishing gains in newer models
• 00:24:23 – Context Engineering
• 00:35:01 – Multi-agent systems and diversity of models
• 00:36:56 – Video Course Spotlight
• 00:38:34 – Current generation of coding agents
• 00:44:00 – Fast vs deep reasoning
• 00:45:18 – Agent Context Protocol
• 00:50:19 – Working through the hype cycle
• 00:55:43 – Open-source contribution pollution
• 00:57:21 – Local models
• 00:58:36 – Rick Beato comparing how the music industry failed
• 01:08:41 – LLMs are an amazing development
• 01:11:33 – Keynote talk on AI summers and winters
• 01:12:45 – PyCon US and EuroPython
• 01:14:11 – Thanks and goodbye

Show Links:

• AI Agent Course - Build a Language‑Learning Agent with OpenAI, LangGraph, Ollama & MCP - YouTube
• Episode #264: Large Language Models on the Edge of the Scaling Laws
• Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs
• Reinforcement learning with verifiable rewards (RLVR)
• What is test-time compute and how to scale it?
• Overfitting - Wikipedia
• 2025 LLM Year in Review - karpathy
• Animals vs Ghosts - karpathy
• Agent Context Protocols Enhance Collective Inference
• Open source AI we use to work on Wagtail - Wagtail CMS
• LLMs for Devs: Model Selection, Hallucinations, Agents, AGI – Jodie Burchell - The Marco Show
• Keynote - Can you trust your (large language) model? - Standard error
• The Human-in-the-Loop is Tired
• How AI Will Fail Like The Music Industry - YouTube
• “Yes, AI Is a Bubble. There Is No Question.” - The Ringer
• Keynote: AI is having its moment … again - Jodie Burchell - NDC Copenhagen 2025
• PyCon US 2026
• EuroPython 2026 - July 13th-19th 2026 - Kraków, Poland
• Jodie Burchell (@t-redactyl.bsky.social) — Bluesky
• Standard error

Level up your Python skills with our expert-led courses:

• Vector Databases and Embeddings With ChromaDB
• Getting Started With Claude Code
• Getting Started With Google Gemini CLI

Support the podcast & join our community of Pythonistas

Efe Gümüs: The People-Pleasing Product Owner and the PO Who Understood User Value — Two Sides of Product Ownership

In this episode, we refer to the SPIDR slicing method.

The Great Product Owner: The PO Who Understood User Value

Read the full Show Notes and search through the world's largest audio library on Agile and Scrum directly on the Scrum Master Toolbox Podcast website: http://bit.ly/SMTP_ShowNotes.

"If your product owner can phrase what the user wants to do — not what the button should look like — it is going to be a night and day difference." - Efe Gümüs

Efe describes the great product owner as someone who creates focus and a clear product vision, so the team knows what they're building and why. The foundation is simple but powerful: describe what the user will be able to do, not what the interface should look like. Instead of specifying a red subscribe button with exact text in three languages, say "as a user, I want to subscribe to my favorite channel." That shift unlocks the team's ability to contribute design insights, architecture decisions, and user journey thinking — the kind of expertise no product owner could anticipate alone. Efe highlights the SPIDR slicing method as one of his favorite tools for breaking product backlog items into consumable pieces — by interface (iOS, Android, web), by data, by rules. When the PO frames work around user value and slices it effectively, the team delivers visible value in iterations, and sprint goals become meaningful. Without this, the team becomes a ticket delivery machine.

Self-reflection Question: When you look at your product backlog right now, are items described in terms of what users can do — or in terms of what the interface should look like?

The Bad Product Owner: The People-Pleasing PO Who Says Yes to Everything

Read the full Show Notes and search through the world's largest audio library on Agile and Scrum directly on the Scrum Master Toolbox Podcast website: http://bit.ly/SMTP_ShowNotes.

"If you are doing everything your customer says, then you are not managing your product. That's the foundation." - Efe Gümüs

Efe names people-pleasing as the worst product owner anti-pattern — the "customer is always right" mentality applied to product management. When a PO says yes to every request, the consequences cascade quickly: multiple priorities competing simultaneously, everything marked urgent, no meaningful sprint goal, constant context switching, and new items injected mid-sprint. The team loses focus entirely. Efe has seen this in startups where the CEO walks in with urgent customer requests, and in larger organizations where multiple customers each demand customizations. In both cases, the PO becomes a pass-through instead of a decision-maker. The customer might be happy today, but will they be satisfied in six months when nothing is coherent? As Vasco notes, when you're serving multiple customers and saying yes to one, you're saying no to all the others — you just haven't told them yet. The result is chaos: steering blindfolded without navigational tools, trying to go everywhere at the same time. A product owner's most important skill is coherent, aligned decision-making — and that means learning to say no.

Self-reflection Question: How often does your product owner say no to stakeholder requests — and when they do say yes, is it because the request aligns with the product vision or because they want to avoid conflict?

[The Scrum Master Toolbox Podcast Recommends]

🔥In the ruthless world of fintech, success isn't just about innovation—it's about coaching!🔥

Angela thought she was just there to coach a team. But now, she's caught in the middle of a corporate espionage drama that could make or break the future of digital banking. Can she help the team regain their mojo and outwit their rivals, or will the competition crush their ambitions? As alliances shift and the pressure builds, one thing becomes clear: this isn't just about the product—it's about the people.

🚨 Will Angela's coaching be enough? Find out in Shift: From Product to People—the gripping story of high-stakes innovation and corporate intrigue.

Buy Now on Amazon

[The Scrum Master Toolbox Podcast Recommends]

About Efe Gümüs

Efe is an out-of-the-box Agile Coach and Scrum Master who brings fresh perspectives to Agile by connecting it with everyday life. He uses metaphors to reveal mindset patterns and applies continuous feedback loops beyond work, including music production and gym training, constantly refining performance, creativity, and personal growth and resilience.

You can link with Efe Gümüs on LinkedIn.

About Erin

Erin Coupe is an international keynote speaker, executive advisor, and bestselling author who advises high-performers and organizations on redefining how success is achieved and sustained. Her work equips leaders to strengthen clarity, expand capacity, and sustain high performance over time, enabling them to operate with greater discernment, presence, and effectiveness in today’s most demanding environments.

She is the author of the bestselling book I Can Fit That In, named a 2026 NextList selection by J.P. Morgan, recognizing influential ideas shaping the future of leadership, culture, and performance. Her insights have been featured in Fast Company, Inc., Business Insider, Success Magazine, The Guardian, and leading media outlets. She is a trusted advisor to executives and leadership teams across Fortune 100 to Fortune 1000 companies, global associations, universities, and professional sports organizations.

Erin is known for her rare ability to articulate the invisible pressures shaping modern leadership and translate them into practical, transformative shifts. Through keynote speaking, executive advisory, and immersive leadership experiences, she equips individuals and organizations to expand leadership capacity, strengthen clarity, and lead with grounded authority rather than constant urgency.

Prior to starting her own business, Erin spent over 17 years inside publicly traded corporations, including Goldman Sachs, where she experienced firsthand the pace, expectations, and complexity leaders navigate at the highest levels. This lived experience, combined with her expertise in neuroscience, human behavior, and leadership development, allows her to bridge the gap between high performance and human sustainability in a way that resonates deeply with executive audiences.

She lives on the North Shore of Chicago with her husband and two boys, where life beyond work serves as her source of ritual and perspective.

Today We Talked About

• Erin’s background
• Turning Pain into Purpose
• Bad Leadership
• Put your oxygen mask on first
• Who do you want to model
• Healthy Boundries
• Leadership
• Fitting in the important things
• Invisible Loads…
• Outdated beliefs
• Always business

Connect with Erin

• LinkedIn
• Website
• Book

Leave me a tip $
Click here to Donate to the show

I hope you enjoyed this show, please head over to Apple Podcasts and subscribe and leave me a rating and review, even one sentence will help spread the word.  Thanks again!

Welcome to the Cloud Wars Minute — your daily cloud news and commentary show. Each episode provides insights and perspectives around the “reimagination machine” that is the cloud.

In today’s Cloud Wars Minute, I analyze how Microsoft is addressing the limitations of pure AI autonomy in enterprise automation.

Highlights

00:09 — Microsoft has introduced new capabilities to Copilot Studio to enhance automated operations by combining AI agents and workflows. Currently, Copilot Studio users can choose between agents and workflows to create automations.

00:27 — While agents are inherently flexible when it comes to business use cases, Microsoft has recognized that, as they state, pure agent autonomy doesn’t always hold up to production requirements. On the other hand, workflows, which are more rigid and rule-based, can be inflexible and may have limitations in their capabilities.

Community Summit North America is the largest independent innovation, education, and training event for Microsoft business applications delivered by Expert Users, Microsoft Leaders, MVPs, and Partners. Register now to attend Community Summit in Nashville, TN from October 11-15.

00:58 — The first pattern involves workflows calling agents to make judgment calls on structured automations. To support this, Microsoft is introducing agent nodes. This allows users to call an existing agent from a workflow, send a message to the agent, retrieve the agent’s response, and use it in subsequent workflow steps if necessary.

01:31 — Now, the second pattern that Microsoft has identified is using workflows as tools. In this scenario, when an agent is working through a complex task, instead of trying to learn how to handle it independently, it can call an existing workflow to execute the subprocess and then continue its reasoning based on the results.

02:29 — Microsoft states that these two approaches combine agents and workflows to provide users with the flexibility to build automations that better address real-world needs—and that’s the key here: real-world applications.

The post Microsoft Introduces Hybrid AI Automation in Copilot Studio appeared first on Cloud Wars.

The following article originally appeared on Jennifer Pahlka’s Eating Policy website and is being republished here with the author’s permission.

I read Norman Maclean’s Young Men and Fire when I was a teenager, I think, so it’s been many years, but I still remember its turning point vividly. It’s set in 1949 in Montana, at the Gates of the Mountains Wilderness, about an hour north of Helena. A fire is burning, and the Forest Service sends out their smokejumpers to fight it. But the fire changes direction without warning, and a group of smokejumpers working in the Mann Gulch find themselves trapped, facing certain death. Instead of running, the foreman, Wag Dodge, pulls out matches and does the unthinkable: He lights a fire.

Today we know what he was doing. The escape fire consumed the fuel around him, allowing the main fire to pass over him and a few of his colleagues. But in 1949, the families of the 13 other smokejumpers who died accused Wag of causing their deaths. To them, what he had done made no sense.

I love that Marina Nitze, Matthew Weaver, and Mikey Dickerson chose this story as a framing device for their new book, Crisis Engineering: Time-Tested Tools for Turning Chaos Into Clarity, out now. Not just because it brought back the memory of a book that I once loved, but because Maclean’s obsessive investigation of what had happened back then (he wrote the book years after the incident) seemed to me almost as heroic as the bravery of the smokejumpers. And indeed, his insistence on making sense of what happened has probably saved lives. Escape fires are now formally recognized and taught as a last resort tactic when training new firefighters.

The Dodge escape fire wouldn’t seem to have much to do with Three Mile Island or healthcare.gov or the pandemic unemployment insurance backlogs, but the authors use it to make a point about how action and understanding interact in a crisis. One key is exactly what Maclean himself did so well: sensemaking. In a crisis like Mann Gulch, sensemaking disintegrates: a broken radio, wind so strong communication is impossible, fire whose behavior violates well-tested assumptions, and a team scattered. You don’t achieve sensemaking by staring at a map; you achieve it by acting and observing results. Wag Dodge didn’t understand fire behavior well enough to explain the escape fire in advance. But his actions created the understanding itself—retrospectively, as all real sensemaking is.

The book’s key claim is that crises are opportunities, and the authors leverage Daniel Kahneman’s Thinking, Fast and Slow to explain why crises are the only real windows for organizational change—and why everything else, the incentives, the logical arguments, the reorganizations, mostly doesn’t work. Most organizations, most of the time, run on autopilot. People habituate to their environment, rationalize away small surprises, and build stable stories about how things work. A crisis breaks this. When surprise accumulates faster than the brain’s “surprise-removing machinery” can rationalize it away, the whole apparatus jams, and organizations become, briefly, reprogrammable.

An institution resolves a crisis in one of three ways, according to the authors. It makes durable deliberate change, it dies, or, most commonly, it rationalizes the failure into an accepted new normal. “Most large organizations contain programs and departments that passively accept abject failure: infinitely long backlogs, hospitals that kill patients, devastating school closures that do little to affect a pandemic. These are fossils of past crises where the organization failed to adapt.”

Too many of our public institutions have failed to adapt, and the idea that they might be reprogrammable at all is a bit radical. We live in an era when too many people have given up on them, willing to burn them to the ground rather than renovate them. If crises represent the chance for true transformation, then we’d better get a lot better at using them for that. This is explicitly why Crisis Engineering exists, and it’s a detailed, practical book—the theory and framing devices are well used, but there’s a ton of pragmatic substance here you’ll be grateful for when the moment comes.

I remember when I was working in the White House and frustrated by the slow pace of progress. My UK mentor Mike Bracken told me: “Hold on, you just need a crisis. You Americans only ever change in crisis.” Boom. About two months later, healthcare.gov had its inauspicious start. And he was right. Change followed. Not all the change we needed, but a start. Marina, Weaver, and Mikey are three of the people who drove that change. I got to work with them again the first summer of the pandemic on California’s unemployment insurance claims backlog. I’m not a crisis engineer, but their strategies and tactics have deeply influenced how I think about the work I do and how I think we’re going to get from the institutions we have today to the ones we need.

We may be living in an era when too many people have given up on institutions, but we are also likely entering an era of crisis, and even polycrisis. This makes for uncomfortable math, but also drives home the need for a new generation of crisis engineers.

When I first read about Mann Gulch, so many years ago, I remember being in awe of the ingenuity and courage it took to start Wag Dodge’s escape fire. Today I think a lot about that pattern: the controlled burns that reduce the risk of megafires, the little earthquakes that take the pressure off faults under great tension, the managed crises that, if we’re skilled enough to use them, keep our institutions from the kind of collapse that comes when nothing has been allowed to give for too long. Dodge didn’t burn things down. He burned a path through. We’re going to have to get good at that.