The Reasoning Show

SUMMARY: How software development is rapidly evolving in the age of AI and automation. Matt Moore shares how his team is rethinking secure software supply chains, scaling infrastructure, and safely integrating AI agents into development workflows.
GUEST: Matt Moore, CTO at Chainguard 
SHOW: 1022
SHOW TRANSCRIPT: The Reasoning Show #1022 Transcript
SHOW VIDEO: https://youtu.be/9Q0kWkTYRs8
SHOW SPONSORS:
ShareGate - ShareGate Protect. Microsoft 365 Governance, we got this!
Nasuni - Activate your data for AI and request a demo
SHOW NOTES:
Chainguard Factory 2.0
DriftlessAF

Scaling Challenges & “Factory” Evolution
Early automation relied on tools like GitHub Actions
At scale, simple systems broke due to:Massive event volumes
API rate limits (e.g., GitHub quotas)
Exponential fan-out effects

Key innovation: custom work queue + reconciliation model~90% event deduplication
Controlled throughput and backpressure
Improved reliability and system stability

Introduced Driftless 
Built on reconciliation principles (inspired by Kubernetes):Compare desired vs. actual state
Continuously reconcile differences

Benefits:Resilience to missed events
Automatic retries and recovery
Scales better than purely event-driven systems

AI Agents in Software Development
AI is dramatically accelerating development workflows
Chainguard uses agents to:Remediate vulnerabilities (CVEs)
Update dependencies
Fix failing tests and adapt to upstream changes

Key Design Philosophy
Least privilege → “least tool call”Avoid giving agents full system access
Provide narrowly scoped tools for specific tasks

Delegate execution to sandboxed systems (e.g., CI pipelines)
Focus on safe, controlled automation
Industry Shift: Velocity vs. Security
Explosion of AI-driven tools (e.g., autonomous PR generation)
Massive increase in development velocity
New risks:Poorly secured agent frameworks
Malicious or unsafe automation patterns

Key Takeaways
Scale changes everythingSimple systems break under massive workloads
Purpose-built infrastructure becomes necessary

Reconciliation > pure event-driven systems at scaleMore resilient, predictable, and controllable

AI is a force multiplier—but requires guardrailsUnrestricted agents introduce serious risk
Constrained, purpose-built agents are safer and more effective

Continuous learning is mandatoryAI tooling is evolving too fast for static skillsets
Teams must actively experiment and adapt

FEEDBACK?
Email: show @ reasoning dot show
Bluesky: @reasoningshow.bsky.social
Twitter/X: @ReasoningShow
Instagram: @reasoningshow
TikTok: @reasoningshow

SUMMARY: How real-time power flow optimization at the edge is helping data centers and the electrical grid handle surging AI energy demands more efficiently. By unlocking hidden capacity and dynamically managing power systems, we explain how existing infrastructure can support significantly more compute without massive new buildouts.
GUEST: Marissa Hummon, CTO Utilidata
SHOW: 1021
SHOW TRANSCRIPT: The Reasoning Show #1021 Transcript
SHOW VIDEO: https://youtu.be/ItcpU8UjOFE
SHOW SPONSORS:
Nasuni - Activate your data for AI and request a demo
ShareGate - ShareGate Protect. Microsoft 365 Governance, we got this!
SHOW NOTES:
Utilidata (homepage)
AI Data Center to Receive 50% Capacity Boost with AI Power Orchestration
KEY TOPICS:
Differences between grid power dynamics vs. AI workloads
Edge AI for real-time power flow optimization
Unlocking stranded capacity in existing infrastructure
“4-to-make-3” vs. “4-to-make-4” data center design
AI training vs. inference power consumption patterns
Role of NVIDIA-powered edge compute modules
Grid modernization and coordination with utilities
Security and resilience in critical infrastructure
KEY MOMENTS:
From centralized AI models to edge-based decision-making
Defining efficiency: utilization vs. thermal performance
Why AI workloads aren’t as constant as they seem
NVIDIA partnership and edge compute in power systems
Using redundancy to increase usable capacity
Increasing density of AI compute and hidden capacity
Data center vs. utility responsibilities
Addressing data center bottlenecks and scaling challenges
Customer landscape: hyperscalers to enterprise
Security, resilience, and critical infrastructure
KEY INSIGHTS:
AI workloads are dynamic, not constant: Training and inference create fluctuating power demands that can be optimized.
Edge intelligence is critical: Real-time sensing and decision-making at the edge unlock efficiency gains not possible with centralized models.
Hidden capacity exists: Many data centers have up to 2x unused power capacity due to lack of visibility and control.
Software-defined power is the future: Faster control loops allow systems to safely exceed traditional design limits.
Efficiency = utilization: The biggest gains come from better use of existing infrastructure, not just improving hardware efficiency.
TAKEAWAYS:
AI infrastructure growth is as much an energy challenge as a compute challenge
Real-time, edge-based control systems are key to scaling sustainably
Existing grid and data center investments can go further with smarter orchestration
The future of AI scaling depends on aligning compute innovation with energy intelligence
FEEDBACK?
Email: show @ reasoning dot show
Bluesky: @reasoningshow.bsky.social
Twitter/X: @ReasoningShow
Instagram: @reasoningshow
TikTok: @reasoningshow

SUMMARY: Shadow AI is growing much faster than known AI adoption across businesses. How can IT teams get Shadow AI under control?
GUEST: Uri Haramati, CEO at Torii
SHOW: 1020
SHOW TRANSCRIPT: The Reasoning Show #1020 Transcript
SHOW VIDEO: https://youtu.be/AUrh_xICPzM
SHOW SPONSORS:
ShareGate - ShareGate Protect. Microsoft 365 Governance, we got this!
Nasuni - Activate your data for AI and request a demo
SHOW NOTES:
Torii (homepage)

Topic 1 - Welcome to the show. Tell us about your background and your focus at Torii. 
Topic 2 - Is Shadow AI really a security problem—or is it a product-market fit problem inside the enterprise?
Topic 3 - Why does Shadow AI spread faster—and become more dangerous—than traditional Shadow IT?
Topic 4 - What’s the first signal a company should look for to know Shadow AI is already happening?
Topic 5 - How do you balance visibility vs. control without killing the productivity gains that drove Shadow AI in the first place?
Topic 6 - How should organizations rethink ‘data loss prevention’ in a world where the leak is a prompt, not a file?
Topic 7 - What does a ‘well-governed’ AI environment actually look like in practice—day-to-day for an employee?
Topic 8 - “Do you think Shadow AI ever fully goes away—or does it become a permanent operating model that companies need to design around?”
FEEDBACK?
Email: show @ reasoning dot show
Bluesky: @reasoningshow.bsky.social
Twitter/X: @ReasoningShow
Instagram: @reasoningshow
TikTok: @reasoningshow

SUMMARY: Have we reached a point where coding is a solved problem? And if so, what are the downstream effects on companies that need software to differentiate their business?
GUEST: Brandon Whichard, Co-Host of Software Defined Talk
SHOW: 1019
SHOW TRANSCRIPT: The Reasoning Show #1019 Transcript
SHOW VIDEO: https://youtu.be/q0mksIKcBzk
SHOW SPONSORS:
ShareGate - ShareGate Protect. Microsoft 365 Governance, we got this!
Nasuni - Activate your data for AI and request a demo
SHOW NOTES:
The New Kingmakers (Stephen O’Grady - 2014)
Developer Growth Rates
[Via ChatGPT]  A useful way to think about it:
Typing code → mostly commoditized
Designing systems → partially assisted
Owning outcomes → still very human
Topic 1 - How many years into Public Cloud did we assume that Cloud had solved the IT problem? 
Topic 2 - Developers - what are we solving for?
10% of time coding, mostly on the last 10-15% 
Lots of time in planning meetings (decoding requirements, resource planning, updates, etc.)
Decent amount of time fixing, troubleshooting, technical debt reduction
Topic 2a - Business people have unlimited ideas, and most ideas are money + tech
What would be their interface to problem solving without developers? (is this just a shift to consultants)
Is this a massive opportunity for a great PaaS 3.0 company (e.g. is Vercel an example?)
Topic 3 - [Hypothetical] Let’s assume a fairly normal company fired all their software developers tomorrow. How long before they could get a moderately complex new application of integration into production? 
Topic 4 - Nobody likes to work on legacy code - missing source, missing engineers, etc. What do we call any code written by AI that was abandoned within the last 6-12 months? 
FEEDBACK?
Email: show @ reasoning dot show
Bluesky: @reasoningshow.bsky.social
Twitter/X: @ReasoningShow
Instagram: @reasoningshow
TikTok: @reasoningshow

SUMMARY: The RAG (Retrieval Augmented Generation) pattern is one of the most frequently used to augment LLMs with context-specific information. Let’s explore RAG. 
GUEST: Roie Schwaber-Cohen, Head of Developer Relations at Pinecone
SHOW: 1018
SHOW TRANSCRIPT: The Reasoning Show #1018 Transcript
SHOW VIDEO: https://youtu.be/-kZZEMR341Q
SHOW SPONSORS:
Nasuni - Activate your data for AI and request a demo
ShareGate - ShareGate Protect. Microsoft 365 Governance, we got this!
SHOW NOTES:
Topic 1 - Welcome to the show. Tell us a little bit about your background, and what you focus on these days at Pinecone 
Topic 2 - Let’s begin by talking about RAG systems. What are they? Why do companies choose to use them? What benefits do they provide in AI systems?
Topic 3 - At a high level, RAG sounds straightforward—retrieve relevant context, generate an answer. But in practice, where does it break first as systems scale?
Topic 4 - I’ve heard that RAG systems can return answers that are technically correct but fundamentally wrong. What’s a concrete example of that happening in production—and why does it slip past most teams?
Topic 5 - In traditional systems, we assume there’s a single source of truth. But in enterprise environments, ‘truth’ is often versioned, contextual, and conflicting. How should teams rethink ‘truth’ when building AI systems?
Topic 6 - A lot of teams assume their knowledge base is ‘good enough’ for RAG. What do they usually underestimate about the messiness of real enterprise data?
Topic 7 - There’s a growing narrative that better reasoning models can compensate for weaker retrieval. From what you’ve seen, where does that idea fall apart?
Topic 8 - If correctness depends on things like timing, policy scope, or configuration, how should teams design systems that understand context—not just content?
Topic 9 - Looking ahead, what replaces today’s RAG architectures? What patterns are emerging among teams that are actually getting this right?”

FEEDBACK?
Email: show @ reasoning dot show
Bluesky: @reasoningshow.bsky.social
Twitter/X: @ReasoningShow
Instagram: @reasoningshow
TikTok: @reasoningshow