2025 Key Trends: AI Workflows, Architectural Complexity, Sociotechnical Systems & Platform Products

In this end-of-year InfoQ podcast panel, industry leaders reflect on AI's transformative impact on software delivery, the critical role of managing architectural complexity, the evolution of sociotechnical systems, and the maturation of platform engineering. Key takeaways include AI amplifying organizational dysfunction, the necessity of clean architectural boundaries even with AI, concerns over ethics and sustainability, and a shift towards treating platforms as products.

Introduction

In this end-of-year panel, the InfoQ podcast hosts reflect on AI’s impact on software delivery, the growing importance of sociotechnical systems, evolving cloud realities, and what 2026 may bring.

Key Takeaways

AI is now reshaping how software gets built. We've moved from "AI is impressive" to "AI is changing workflows": agentic systems, MCP and Agent2Agent-style interoperability, and AI becoming more like a team member (task prioritisation, calendar/workflow integration, co-creation of software).
AI accelerates everything, including (potentially) your org’s existing dysfunction. Strong teams with good engineering practices get better/faster, while weaker teams get more chaotic. AI doesn’t magically fix mediocre DX, culture, or delivery pipelines: it amplifies friction points.
Managing complexity is the architect’s core job, and AI raises the stakes. There’s explicit resistance to the idea that "AI can handle complexity so humans don’t need clean boundaries". Instead: separation of concerns, DDD, smaller components, and clear intent matter even more when AI is generating (and accelerating) change.
Alongside excitement, there’s concern about AI: ethical focus slipping, sustainability being treated as cost-only, burnout/"996" pressures, and juniors’ development in an AI-saturated workplace. Team trust is highlighted as fragile if orgs treat people as interchangeable because "the AI can do it".
Platforms, cloud, and reliability are back in focus. "Global by default" is fading, as seen in outages and multi-region reality checks, as well as Europe’s growing sovereignty/regional concerns. On platform engineering specifically, there’s a "trough of disillusionment" vibe, plus a renewed push toward better abstraction/composition layers and treating platforms as a product.

Panel Introductions

Daniel Bryant: Welcome to the InfoQ Podcast. We have a fantastic panel of folks, leaders within the InfoQ space. We'll start with introductions.

Renato Losio: I'm a cloud architect at Cloudiamo, based in Berlin. I'm an editor for the Cloud and DevOps Queue, and my highlight has been running the InfoQ Dev Summit in Munich.

Srini Penchikala: I work as an enterprise architect and serve as the lead editor for the AI, ML, and Data Engineering Group. I'm currently serving as a program committee member for QCon London 2026.

Thomas Betts: Thomas Betts, working at Blackbaud and doing a lot of QCon stuff. I'm accepting a new position as the architect for our new Agents for Good lineups, creating AI agents for the social good community.

Shane Hastie: I'm Shane Hastie, lead editor for Culture & Methods. Still with Skills Development Group as my day job, deeply entrenched in the human side of sociotechnical systems.

Daniel Bryant: I'm still at Syntasso, working on platform engineering. We're doing a lot of stuff with the sociotechnical side of platform building.

Reviewing 2024/2025 Predictions

Daniel Bryant: Let's look back at our predictions from last year.

Thomas Betts: I talked about sustainability. I thought architects would start considering it as a factor, but with the shift to AI, I thought it'd get lost. It's actually sneaking back in because customers are asking about the environmental impact of AI. So, I'm half right for different reasons.

Shane Hastie: I predicted getting past knee-jerk return-to-office mandates. I'd say I've been half right. Some organizations are still doing it, others have faded. I also predicted the AI partner would be huge but risk losing human critical thinking. Some studies indicate that is happening, which is a worry.

Renato Losio: I said last year that I expected Intel not to be the default processor. We're probably still far away from that. I also said we were going away from regional endpoints to global. I got that entirely wrong due to the political landscape; in Europe, the focus is on separate cloud regions and sovereignty.

Srini Penchikala: I mentioned that we would arrive in AI adoption when we don't have to talk about AI as a thing anymore. We aren't there yet, but we are closer with innovations like AI agents and protocols like MCP and Agent2Agent.

Daniel Bryant: I said platform engineering would head into the trough of disillusionment. I'm seeing a lot of folks struggling with the day-two story and upgrades. I also predicted better abstractions; we're seeing frameworks like kro, Crossplane, and Kratix emerge, but it's a work in progress.

Managing Complexity in the Age of AI

Daniel Bryant: How are architects managing complexity?

Thomas Betts: The one-sentence definition of an architect is they manage complexity. That's both explaining it to people (the socio part) and encapsulating it so we don't have a big ball of mud. DDD, separation of concerns—these help keep cognitive load minimal. With AI, we have to consider how it affects team dynamics. I resist the idea that AI doesn't need separation of concerns; we still need human-understandable pieces.

Shane Hastie: The AI as a team member is a key theme. The danger is abdicating thinking to AI tools. One study showed 300% more code but 400% more bugs. We've spent decades aiming for smaller components (microservices), and AI can produce a monolith nobody understands if you aren't careful. Prompt engineering and training the bots are new skills.

Renato Losio: In small projects, AI helps me go faster, but I accept extra complexity in the code. Is that ideal? No. Where do you find a balance? I have no idea. For larger projects, I don't know if you can scale that.

Thomas Betts: It encourages us not to use lines of code as a metric. Nicole Forsgren said AI amplifies friction. If your developer experience has friction, AI will rub against it 100 times faster. We're now working at the speed of computers, not people.

Shane Hastie: The DORA report is shifting to the state of AI-assisted software engineering. Good engineering teams with AI support get better, but average teams get more average. There's a bifurcation happening.

Thomas Betts: Culture eats strategy for breakfast. High-performing teams have a growth mindset; they see AI as a tool. Fixed mindset teams try to apply AI without changing how they work.

Shane Hastie: Organizations mandate AI without thinking about implications. "Everyone else is using it, so we must too."

The Impact of Agentic AI

Thomas Betts: At QCon SF 2024, agentic AI was future-looking. Now, it's here. MCP (Model Context Protocol) wasn't on our trends report six months ago; now everyone talks about it. If you haven't optimized for this new way of working, you'll trip over yourself. Smaller, focused agents are better than large ones. We're seeing old patterns like the Actor pattern become mainstream.

Renato Losio: When we discussed the Munich keynote in March, we debated whether to explain what MCP was. By September, we removed the description because it was obvious. That's how fast things change.

Srini Penchikala: AI as a team member is emerging. Companies like Glean envision AI integrating into workflows to alert team members on tasks and manage calendars. Martin Fowler talks about "expert generalists"—people with a blend of generalist and specialist skills. This is even more relevant with AI adoption. As Shane said, AI won't replace people, but people who use AI better will. Companies that don't adapt will be replaced by those who do. However, some companies are going overboard with AI-first mandates where it's not a good fit.

Daniel Bryant: Big plus one to Martin Fowler's "expert generalist" theme. It's fantastic.

InfoQ AI and ML Trends Report Highlights

Srini Penchikala: 2025 was the year AI was talked about most. Highlights from our report:

Language Model Innovations: Vision language models, small language models (SLMs), and reasoning models (like GPT-5 thinking) are emerging. Small models are gaining traction for on-device inference and privacy.
Infrastructure: Tools like vLLM and LLM-D help teams run LLMs on-premise, avoiding expensive cloud solutions and privacy concerns.

Ethical Concerns and Sustainability

Daniel Bryant: Shane, you wanted to discuss ethics.

Shane Hastie: A year ago, there was stronger conversation about ethical implications. Now, organizations are slipping away from ethical guidance. There's less concern about "just because we can, doesn't mean we should." Sustainability is often linked only to cost reduction, not environmental or human impact. We see a return to "996" culture (working 99 hours a week, 6 days a week), leading to burnout. People are scared, and it's not sustainable.

Daniel Bryant: I've seen pushback in the Bay Area, but it's a challenge.

Renato Losio: I'm more pessimistic. We use sustainability as a proxy for cost reduction. VCs push back on burnout only because they don't see results, not because they care about ethics.

Cloud Computing Stories in 2025

Daniel Bryant: Renato, what are your cloud highlights?

Renato Losio: The main story of Q4 was outages (e.g., AWS Northern Virginia). These outages were good for the industry because they showed how much infrastructure depends on the cloud. Even if you claim to be multi-region, many got lucky rather than being truly resilient. The reaction was a return to "move back to your own data center," but I believe for most workloads, the cloud isn't the weakest link—your own infrastructure is. In terms of providers, Cloudflare is interesting because they are extending their platform to be developer-focused. At re:Invent, there was a disconnect between AI announcements and what developers actually wanted. Durable Functions and serverless running on your own infrastructure were side announcements against the AI hype.

Thomas Betts: Every company is sticking "AI" on everything. Customers are asking, "Could you just improve the software?" AI is becoming a feature people don't want but will eventually expect. We aren't at the magic wand stage yet.

Renato Losio: Cloud providers used to be trusted to keep services around. Now, with AI, there are 200 announcements, and by the end of the year, half are gone.

Platform Engineering Evolution

Daniel Bryant: In platform engineering, chat interfaces are succeeding in observability (e.g., Incident.io, Dash0), allowing drill-downs and synthesis. Security hasn't found the right AI interface yet. There's a push toward building platforms as a product—cohesive ecosystems, not just collections of tools. Susanne Kaiser's new book combines Team Topologies, DDD, and Wardley Mapping to address the sociotechnical aspects of building platforms.

Shane Hastie: We're seeing dynamic team structures adopting AI as a teammate. Role definitions are blurring (e.g., product managers producing code). Feedback cycles are shorter. The concern is: what about juniors coming into the team? How do we ramp them up when team dynamics are shifting so fast?

Thomas Betts: Highly effective teams rely on trust. With AI, there's a risk of viewing people as interchangeable cogs. If your only value is writing code, AI can replace you. The value is in solving problems and interacting with people. We need to keep trusting each other despite changing dynamics.

Predictions for 2026

Renato Losio: By the end of 2026, we will have forgotten 80% of the current AI announcements. A major shift in open source is happening: cloud providers are collaborating to fork projects (like Valkey, DocumentDB) and donate them to the Linux Foundation, taking ownership away from original creators who changed licenses.

Thomas Betts: MCP will be everywhere. AI will become a first-class design principle (architecting for AI). However, I predict the AI bubble might burst or contract in 2026 due to hype vs. reality.

Shane Hastie: The bifurcation will continue—good teams get better, bad teams get worse. Sensible organizations will focus more on human skills (critical thinking) and building adaptable team environments.

Srini Penchikala: The next frontier is Physical AI (manufacturing, logistics). RAG becomes a commodity, replaced by Enhanced Agentic RAG (EA RAG). AI shifts from assistant to co-creator of software. AI-driven DevOps will be huge. Voice will become a prominent interface. AI Security will see significant developments with more guardrails.

Daniel Bryant: AI will expose brittle platforms faster than it fixes them. It will rejuvenate investment in platform engineering because folks are struggling to deliver AI projects. However, if your platform is brittle, going 10X faster with AI will cause it to fall apart. We will see big failures where the root cause is a lack of guardrails and a poor sociotechnical setup.

Wrapping Up

Daniel Bryant: Hopefully, this has been food for thought as we kick off the new year. Stay tuned for more podcasts, news, and articles. Thank you for listening, and Happy New Year!