Can an engineering team double its AI usage without burning through token budgets or compromising code security?
Steve Adams, Chief Digital Officer, and Jeff Wooden, Chief Architect at Intevity, believe that navigating the flood of modern AI products requires moving past shiny objects and focusing on practical execution.
As development teams adopt agentic tools and LLM harnesses, organizations quickly run into severe bottlenecks: untracked token spend, fragmented toolkits, and the hidden risks of unvalidated code.
Steve and Jeff talked about:
Steve observed that many organizations rush into AI adoption without clear alignment on business value.
Rather than building commercial products for the sake of hype, Intevity Labs began as an organic internal effort to empower developers, increase engineering velocity, and maintain software quality.
“Jeff and I talked a lot about this idea of how do we transition from the mindset of traditional engineering to more architecture and orchestration,” Steve told our host, Intevity CEO George Jagodzinski. “And then, the second piece was really about how we’re creating the services that align with our clients' needs... As we started to go through solving those problems, part of that was creating tooling. We [addressed] a lot of the challenges with our clients, at scale that a lot of organizations are facing, and so we wanted to share that.”
Jeff highlighted that releasing open-source tools like Sentinel and Trove provides a dual benefit: solving real client friction while creating a safe platform for internal teams to learn and experiment.
“I think we saw gaps in ways we can make the teams more efficient,” Jeff explained. “The other thing that comes out of this is the ability to upskill our teams as well by providing a platform that may not necessarily be client work that they can utilize to learn and grow and put out their own applications.”
While frontier AI models offer unprecedented capabilities, token spend functions like a volatile commodity that can rapidly blow past budgets.
Furthermore, granting local AI agents access to developer environments creates immediate cybersecurity risks if those tools operate without visibility.
To solve these governance and security gaps, Jeff built Sentinel as a local companion tool designed to monitor agent actions and optimize prompt overhead.
By implementing curated sub-agents and response compression, teams can route simpler file operations to lower-cost models while stripping bloated instructions from MCP servers.
“When I first implemented Sentinel, security was top of mind and it would show me things that the agent was doing locally, like reading my SSH keys. Well, that's not cool,” Jeff shared. “I want to know about these things. I want to be able to block certain things if they happen. And so it adds an extra layer of security.”
On the financial side, Sentinel’s optimization layers deliver dramatic efficiencies, as well.
“Overall, I find, even with my regular usage, I'm typically saving around seventy percent of my tokens in a given day, just by having these optimization features on, which is a huge win,” Jeff noted.
As engineering teams scale their concurrent AI workloads, they encounter a productivity paradox: developers feel significantly more productive, yet overall sprint velocity doesn't automatically jump.
Part of this lag comes from shifting estimation baselines, but a larger risk stems from the massive volume of synthetic code entering repositories.
Steve warned of a growing phenomenon known as “dark code”: large volumes of AI-generated code that no human has ever directly inspected or reviewed line-by-line.
When humans lack the bandwidth to review thousands of lines of daily output, relying on full automation for critical systems introduces severe operational risk.
“With that productivity becomes this huge amount of code that humans aren't necessarily touching firsthand, so they have to go through a code review process,” Steve explained. “And then, to a certain extent, there's this black box of code that nobody's ever touched... We call that ‘dark code,’ and it's basically this idea that without having hands-on [review], there's this inherent risk that there's code that's potentially going live or that's in products that nobody's ever touched.”
Steve emphasized that organizations must distinguish between systems of innovation and systems of record. He suggested that winning with AI isn't about automating every workflow, but knowing when to keep human gatekeepers in the loop.
“There's this idea that [if] you move to the right on that scale to full automation, that's winning, but it's not necessarily if the stuff that you're working on is really critical,” Steve said. “Success and what winning looks like isn't necessarily just automating everything.”
Ultimately, navigating the AI transformation requires balancing rapid technological experimentation with grounded human effort.
The core philosophy of Steve's approach to execution can be found in the foundational advice he received on maintaining balance amidst high-powered workflows:
“It was funny when Jeff was talking... I remember this quote: ‘Nobody ever on their deathbed said, “I wish I'd worked more,’” Steve recounted. “So if we can use these tools to give us some balance back, I think that'd be great.”
Craving more? You can find this interview and many more by subscribing to Evolving Industry on Apple Podcasts, on Spotify, or here.