Coding Was Just the Beginning

There's a version of this conversation that stays abstract. AI will transform software engineering. Agentic workflows are coming. Knowledge work will never be the same. Srinivas Narayanan, CTO of B2B Applications at OpenAI, didn't give that version. He gave the one where you find out that a team inside OpenAI just shipped an internal project where a human did not write a single line of code.

"A new project — the repository was 100% created by Codex. A human did not write a single line of code." He was quick to note the nuance: humans reviewed the code, guided the agent, stayed involved throughout. But the code itself was entirely generated. That's not a demo. That's how one of OpenAI's own teams now ships software.

What made the conversation with Bloomberg's Rachel Metz useful wasn't the headline. It was the specificity about how the transformation actually happened — and what it implies for the companies watching from the outside.

From assistants to agents: the shift that happened faster than expected

Narayanan drew a clear line between two eras of AI-assisted coding. A year ago, the pattern was assistive: engineers used AI to fill in specific things, complete specific tasks, move a little faster. Useful, but bounded. Then something changed. "In the last few months, it has become entirely agentic." Every engineer at OpenAI is now running multiple agents in parallel — five, six, seven, ten at a time — guiding them to generate software, review code, fix bugs.

The shift in the job description is the part worth sitting with. "The job of a software engineer has now become more like the job of a CEO." Less about how you accomplish things, more about what you actually want to accomplish. The engineering work is still happening — it's just happening one layer of abstraction higher. You're managing a fleet of agents, not writing the code yourself.

One consequence Narayanan flagged is the blurring of discipline boundaries. Product managers can write code. Designers can write code. Engineers are increasingly doing product management work. "The people with a lot of agency, with a lot of ambition, with a lot of ideas for what they want to build are becoming incredibly more productive." AI isn't just changing what engineers do. It's changing who gets to do what used to require an engineer.

What this looks like at scale outside of OpenAI

The session didn't stay inside OpenAI's walls for long. Narayanan moved quickly to what's happening at customers: Cisco deploying Codex across 18,000 people, Virgin Atlantic compressing website projects from months to weeks, Rakuten using it for debugging at scale.

But the more interesting argument was the one underneath the customer examples. Coding, he argued, is now the template for what's coming in every other domain of knowledge work. "What we are realizing is that the underlying model that is powering coding — a lot of knowledge work has the same primitives. It's about how you manipulate files, how you access files, how you use a computer." The same primitives that make Codex powerful for software development are now being applied to financial services workflows, legal services workflows, any domain where work is fundamentally about manipulating information.

This reframes the stakes of the coding moment. It's not just that software engineering is changing. Coding is proving out a model of agentic work that will propagate across every function in the enterprise. "Coding as the underlying harness for actually creating agentic workflows in many, many different areas."

Three buckets — and where most companies actually are

Metz pushed on what enterprise customers are actually asking for, and Narayanan's answer organized the demand into three categories. The first is individual productivity: AI enabling people to do entirely new things, not just existing things faster. He cited an OpenAI study finding that 70% of workers across various companies said AI had enabled them to do a new job they'd never been able to do before. Designers writing code. Business users answering data questions that previously required a data scientist. The first bucket, in his framing, is already well underway.

The second bucket is organizational transformation: workflows that run more efficiently, customer support being the clearest example. Agents handling the majority of queries, escalating to humans only when genuinely needed. This bucket, he said, is starting to happen.

The third bucket is the hardest and the most valuable — applying AI to core innovation: new product development, new discovery, genuinely new things a company couldn't do before. This one is earlier. But the first two buckets are what get you to the third. Companies that are still treating AI as a productivity layer are, in his framing, working their way toward the question that actually matters.

The safety question nobody wants to skip past

Metz brought up the failure cases that anyone deploying agents has either seen or heard about: agents deleting email inboxes, agents spamming people, agents making decisions outside their intended scope. Narayanan didn't dismiss these. "With any new set of tools and new technology, there's always novel risks. It's extremely important for us to be forthcoming about it."

His framework for thinking about this had several layers. Model safety — how the model is trained, what guardrails are applied during post-training — is one. But agentic systems create new surface area. An agent with access to both a private Gmail account and a public GitHub repository needs to know not to move sensitive information between them in ways a human would instinctively avoid. Teaching models to respect those boundaries at the level of judgment, not just rule-following, is still an open problem.

His response to this was layered: guardrails at the model level, guardrails in the prompting, red teaming to surface failures before deployment. OpenAI's acquisition of Promptfoo was his concrete example — a tool that helps companies do systematic red teaming on their agent implementations, finding failure modes before users do.

He addressed the agent-to-agent question directly too, when Metz raised the scenario of agents colliding with other agents at scale. "We're definitely evolving into a world where there is going to be agents talking to other agents." His read: the same safety principles that apply to individual agents will have to be replicated across the multi-agent layer. MCP — Model Context Protocol — is the early industry convergence around how agents communicate with tools and with each other. But on identity, on trust hierarchies, on what it means for one agent to authorize another: "There is a lot of innovation left to happen here."

What it actually means to manage agents

The conversation closed on a question that cuts against a common anxiety: not everyone wants to be a manager. If the future of knowledge work is managing fleets of AI agents, is that actually a future people want?

Narayanan's answer was direct. "Managing AI will be easier than managing people." And then the more interesting point: the shift isn't really about management at all. It's about what becomes possible when execution stops being the bottleneck. "You can aim higher. If you have ideas, you can get them executed faster." The constraint moves from capability to ambition. The people who will get the most out of this shift aren't necessarily the ones who are best at running agents. They're the ones with the most ideas about what to build.

That's a different anxiety than the one most enterprise AI conversations are organized around. The question isn't whether you have enough engineers. It's whether you have enough people who know what they want to build — and are willing to aim at something bigger than what they could have executed before.

Watch the full session on-demand. 

Keep the conversation going. Register now for HumanX 2026 Amsterdam (September 22–24, 2026) or HumanX 2027 Las Vegas (March 7–10, 2027). Be sure to subscribe to The Output newsletter and save yourself from bad AI takes (and save on your HumanX ticket).