Overview of The Stack Overflow Podcast
In this first half of Leaders of Code, Stack Overflow hosts Ira May and Peter O’Connor speak with Google’s Andi Gutmans about the shift from traditional software development to an “agentic” world where engineers increasingly guide, review, and orchestrate AI agents instead of writing every line themselves. The conversation draws a parallel between PHP’s role in democratizing web development and today’s AI tools, arguing that agents can similarly expand who can build software—while raising the importance of trust, governance, data quality, and human judgment.
Main Themes and Takeaways
Agents are the next step in democratized development
- Gutmans sees AI agents as a natural evolution of the PHP era: both make software creation more accessible to non-experts.
- He cites examples like non-technical users building websites with tools such as Lovable.
- The big difference now is that agents can also help enforce best practices around security and quality.
Engineering value is shifting from coding to judgment
- The core skills that matter are still:
- understanding the business problem
- designing a sound architecture
- ensuring operational excellence and scalability
- What changes is the time allocation:
- less manual coding
- more reviewing, guiding, and orchestrating agents
- Gutmans’ framing: every individual contributor becomes a “team lead of agents.”
Human review still matters, but should be risk-based
- Google uses both human and agent code review.
- Gutmans argues that review should be applied based on risk:
- highly sensitive/security-related code should get human expert review
- lower-risk changes, like CSS/HTML, may be better handled by agents first
- He emphasizes that the real question is not whether AI is perfect, but where human oversight creates the most value.
AI changes how engineers are assessed
- The discussion suggests interviews should focus less on quick coding drills and more on:
- reasoning
- system design
- how candidates would orchestrate agents
- Gutmans says Google is moving toward interview processes where candidates use Gemini/agents to solve problems, rather than writing everything by hand.
- This highlights aptitude, learning ability, and problem-solving over memorized syntax.
Data, not models, is the biggest bottleneck
- Gutmans believes model capability is already strong enough to automate a large share of enterprise workflows.
- The harder problem is making all enterprise data AI-ready:
- structured data
- operational data
- unstructured data such as PDFs, contracts, and images
- Success depends on understanding the semantics and relationships in the data, not just storing it.
Cross-cloud, secure access is becoming practical
- He describes Google’s work on “borderless” data access:
- data can remain in multiple clouds or on-prem systems
- secure interconnects and fine-grained permissions can still be honored
- The goal is to let agents reason over data wherever it lives, without costly or brittle migrations.
Open formats and interoperability are winning
- The conversation touches on industry pressure toward open data formats like Iceberg.
- Gutmans says customers want zero-copy, cost-effective access to their data across vendors.
- Google’s open knowledge format is positioned as a way to capture knowledge in a portable, reusable way that works across platforms.
Ontologies and knowledge catalogs may become agent-driven
- Gutmans contrasts older ontology systems like OWL, which were powerful but hard to implement, with a newer approach where agents can infer relationships in messy data.
- Google’s Knowledge Catalog aims to shift semantic modeling from human-driven work to agent-driven work.
- Humans will still curate and validate, but agents can do the heavy lifting.
Practical Implications for Engineering Teams
- Adopt risk-based review practices rather than reviewing everything manually.
- Treat agents as teammates that can draft, test, inspect, and critique work.
- Invest in data foundations if you want AI agents to be useful in enterprise settings.
- Modernize hiring and training to emphasize system thinking, reasoning, and orchestration.
- Prioritize open, interoperable data access so agents can operate across silos safely.
Notable Ideas and Quotes
- “Every individual contributor now becomes a team lead of agents.”
- “The value moves toward making sure you’re using your judgment on how to guide agents.”
- “The biggest bottleneck right now is how we make sure data can be activated with AI in the best possible way.”
- The Waymo analogy: even when AI systems are statistically safer, people may still prefer human control—showing that trust and perception matter as much as raw performance.
Episode Context
- This is Part 1 of the conversation.
- The episode ends by teeing up a second half, which will continue the discussion on AI, enterprise data, and the future of engineering work.