AI-led coding using products like Cursor, Windsurf, Claude Code, and GitHub CoPilot has altered how developers write code, yet far too many teams still view AI solely for code generation. This is a very myopic view on both the potential of AI and what we think of developer productivity. In reality, the biggest challenges in software engineering arise not from code creation but from running and maintaining complex production systems. With AI-generated code, this problem is bound to get exponentially worse, and organizations that persist in this narrow perspective risk falling behind.

At the Beyond the Build event in San Francisco, leaders from Salesforce, DoorDash, Replit, DataStax, and Resolve AI convened to explore the challenge and opportunity of agentic AI in software engineering. Their consensus was clear: AI is ready and it’s time we embed it into production environments to address critical operational issues at scale. Engineering leaders must evolve their strategies and recognize that teams neglecting to incorporate AI into their full software engineering lifecycle will be outpaced and out-innovated.

Engineering Leadership Must Push Agentic AI Beyond Code Generation

For many organizations, AI is relegated to generating code snippets, writing unit tests, or automating routine tasks. However, Spiros Xanthos, CEO and Founder of Resolve AI, emphasized that this view misses the bigger picture - "most engineering effort isn’t spent on building new features—it’s consumed by maintaining and troubleshooting in production."

Agentic AI systems are already being used in novel ways to understand production systems and automate essential functions such as incident triage, root cause analysis, and on-call operations. Yet, many senior engineers continue to see AI only as an assistant in writing more code rather than as a transformative force for operational resilience. Amjad Masad, Co-founder and CEO of Replit, emphasized that while many teams treat AI as a tool solely for programming, that perspective is dangerously short-sighted.

"If you're not using AI for programming, you're falling behind. But if you're not embedding AI into your production workflows, your team is setting itself up for failure." — Amjad Masad

In this context, Hari Ramachandran, VP of Engineering at Salesforce and Tableau, shared a bit of frustration about the lack of emphasis of AI on managing operational toil. He went on to elaborate - "I saw a Reddit post that said, 'AI was promised to do all the chores so that I could focus on creativity, but it ended up doing all the cool stuff and I'm doing all the chores for it.' This really hit home. We must stop thinking of AI as merely a coding assistant and start leveraging it to take over the tedious, repetitive tasks that weigh down our production teams."

Together, these perspectives underline an urgent call for engineering leaders: Move past using AI as a productivity tool focused on code generation. Instead, embed it as a core pillar across software engineering: throughout CI/CD pipelines, incident response frameworks, and proactive reliability approaches to drive for real operational excellence.

Agentic AI for Increasing Reliability Within Mission Critical Systems

In today’s cloud-native environment, production reliability is paramount. Shankar Ramaswamy, Head of Engineering at DataStax, an IBM company, detailed that DataStax’s production environment spans over 50 regions across three major public clouds, utilizes thousands of servers, operates 100’s of microservices, and serves tens of thousands of databases, handling between 3 to 5 million requests per second. In such vast and intricate systems, downtime is not a minor inconvenience; it directly impacts revenue and customer trust. From their perspective, if your database is down, your business is down, and in such a business-critical environment, the real opportunity for AI is in automating production reliability.

There is also a mindset shift that leaders must take, that AI doesn’t have to work perfectly right out of the gate. As Shankar put it, “Imagine you're farming with the horse and plow, and now the tractor is here. You'd be an idiot not to take the tractor. Yes, the tractor's not going to work perfectly. There's a million reasons not to do it, but the mindset you want is to see what it can do that you couldn't do before.”

By integrating Resolve AI with incident.io, DataStax has embedded agentic AI into its incident management workflows. This results in significant improvements in Mean Time to Detection (MTTD) and Mean Time to Resolution (MTTR), ensuring that the system operates more reliably and that engineers can respond swiftly to issues.

Scaling Engineering Expertise: Overcoming the Tribal Knowledge Bottleneck

Production systems have become increasingly complex with more fragile dependencies than our teams can track. Our observability tools meant for helping us with this complexity aren’t designed to work well together, creating new fractures. These problems combined, we have a growing volume of signals and data; but the burden of making sense of all the data rests on a handful of highly seasoned engineers.

This creates a critical operational challenge, the persistent reliance on a few senior engineers for troubleshooting, which stifles growth and innovation. Mandar Rahurkar, Director of ML at DoorDash, highlighted how the “same senior engineers get pulled into the toughest incidents repeatedly. If we don't disseminate that expertise across teams, our overall developer productivity will be fundamentally limited."

There is hope, Agentic AI is already showing promise in tackling the problem of tribal knowledge. Agents operating in production systems understand, reason, and learn, the intricate designs of the systems, how the tools work, context of historical incidents, recognize recurring patterns, and are capable of guiding less experienced engineers through resolution processes so that the collective knowledge is shared, reducing the risk of burnout while accelerating recovery.

Where Should Engineering Teams Focus AI Investments

There is also the never-ending debate around whether to build AI systems in-house or to adopt proven external platforms. Hari Ramachandran and Mandar Rahurkar suggested that teams should invest more in customer-facing innovations instead of reinventing existing operational tools.

“What do our users or consumers care about? They care about their order being delivered on time, they care about the experience on the app itself. They don't really care whether we are using Datadog or Resolve AI. So when it comes to tooling, we [would] rather be buying than building some of this ourselves.” - Mandar Rahurkar

The biggest risk of building is to keep up with the ridiculous pace at which AI is evolving. In most cases, by the time you have a team of three people building something, that technology has already been outdated. Or there'll be a better model in a few months.

Key factors to consider in the build vs buy decision:

• Speed of Innovation: External platforms often provide faster access to cutting-edge technology.
• Domain Expertise: Effective AI must grasp the complexities of system architectures and production workflows.
• Data Protection: Ensuring AI systems operate under strict guardrails to protect sensitive information.

Rethinking the Endgame for AI-Driven Software Engineering: Delegation Over Execution

As AI continues to mature, the traditional role of engineers is evolving. Amjad Masad (Replit) envisions a future where developers transition from performing manual tasks to managing intelligent systems.

"The paradigm shift won’t eliminate engineering jobs—it will multiply them. Engineers will lead teams of AI agents that optimize workflows autonomously." — Amjad Masad

AI’s steady migration from “autocomplete for code” to the invisible scaffolding of production systems has surfaced a deeper question: What, exactly, is engineering work once the machines shoulder our routine tasks?

The leaders on stage offered a glimpse:

• Reliability replaces raw output as the true yardstick of craft. When an outage instantly cascades across 50 regions and 5 million requests per second, the only metric that matters is how gracefully the system heals itself.
• Expertise is no longer scarce; it’s atmospheric. When a junior engineer who is on-call can summon an agent versed in every past incident, with a complete understanding of the production system and reason to know how and when to use specific tools, the bottleneck shifts from knowledge access to judgment.
• Stochastic computing unsettles deterministic culture. Engineers raised on binary certainty must learn to manage probability, confidence scores, and imperfect-but-faster answers because a 50 percent-efficient tractor still plows more ground than a perfect horse.

If the last 24 months of AI in software engineering were about shipping code faster, this next wave is about managing complex software production systems. AI in production isn’t a distant aspiration; it’s already emerging as the silent colleague wiring telemetry together, paging itself, conducting robust multi-faceted investigations, and drafting your post-mortem.

In other words, the future of engineering may feel less like typing and more like guiding and tending an ever-learning, ever-shifting ecosystem, one where a developer's value is measured not by how much code they write but their ideas, designs, and by how well they steward a symphony of autonomous agents with humans-in-the-loop humming beneath the surface.