Skip to the stories
Microsoft 365 & SharePoint

Broadcom Pairs Its Young AI Builder With Senior Engineers, Not Unlimited Tokens

Broadcom is pairing a recent graduate's AI security agent with senior engineers and spending heavily on tokens for signature-writing pipelines instead.

Broadcom Pairs Its Young AI Builder With Senior Engineers, Not Unlimited Tokens

Broadcom has taken a different route from Cisco with its young AI talent. Both companies are betting on junior engineers and large token budgets, but Broadcom is attaching a senior safety net to the experiment. My read is that this is less a philosophical split than a difference in how much each company trusts AI-generated output to reach customers unchecked.

Nutanix .NEXT Orb
Broadcom Palo Alto Campus Entrance 2025

Cisco's unlimited-token experiment sets the contrast

The comparison starts with Cisco. A Fortune report in early November described Cisco Chief Product Officer Jeetu Patel revealing that a 27-year-old staffer, nicknamed Paddy, had been given unlimited tokens to rework the company's legacy codebase. That's a lot of latitude for one person, and it made the approach easy to hold up as a headline.

Broadcom's version is less dramatic. Umesh Mahajan, the company's application networking and security chief, told SDxCentral that Broadcom is taking a "more practical" approach. The young hire is paired with senior engineers, partly to stop his agent from going "berserk." Mahajan said he didn't know the specifics of Cisco's program, but he described the rival as "over-marketing it."

Marvell logo

What the Broadcom grad actually built

The project sits inside the company's security team, and Broadcom plans to release it. Mahajan spoke about it during a panel session at VMware Explore in London. He said the graduate built a "wonderful agent" within three months of joining the team.

The agent uses a large language model to translate customers' inconsistent security-rule mappings into a standardised CSV. That file can then be used to generate firewall rules. It's a configuration chore that normally eats more engineering hours than it should, and it's a good example of where an LLM earns its keep: messy, human-written input that needs to become structured data. Broadcom hasn't said when the tool will become a product, so treat this as a capability demo for now.

Groq

Why the senior engineer is in the loop

Mahajan's reasoning is the most useful part of the interview. He said Broadcom typically pairs junior engineers with experienced principal engineers on projects like this. In this case, the graduate works under the company's CTO. "We're letting him be creative, but at the same time the idea is, hey, how do we marry the new approach to the experience and bring out the best of both worlds," Mahajan said.

His concern is about judgement rather than raw ability. "The younger folks, they can be smart. Generally, they know, but ultimately they cannot tell … is this completely right or wrong? They can tell 80, 90%. For that, you need to bring in the absolute domain," he said.

SDxCentral Magazine - Issue 2

That point matters more than it might seem. AI-written code, in Mahajan's words, "will work eight out of 10 cases." The two out of ten are the ones that land in front of customers. Teams in this position typically find that the expensive part isn't generating the code. It's knowing which output is wrong, and that still depends on people who have seen the failure modes before.

Mahajan also said he has told his team not to hire people who can't think analytically. "You don't need 25 years of experience or 10 years of experience," he said. "You can be very good with very little experience. It's how fast your brain works and how you can comprehend and think through."

Google's headquarters in Silicon Valley in Mountain View, California - SDx Cropped
Umesh Mahajan, Vice President and General Manager, Application Networking and Security Division at Broadcom

Where the money is going

Here's the thing: the grad's agent is the small story. The heavier spending is further up the stack. Mahajan told SDxCentral he is prepared to "burn a million dollars" a year on tokens to power an agentic pipeline that writes virtual-patching signatures. The alternative, in his framing, is hiring 150 more engineers or asking existing staff to work longer hours.

The pressure behind that math is vulnerability volume. Broadcom is among the companies given access to Anthropic's Mythos model through Project Glasswing. The company has used the model to unearth a trove of vulnerabilities. Mahajan said his team can currently write 50 to 100 signatures a month, but it needs somewhere between 2,000 and 3,500.

Palo Alto Networks

That gap is the reason he gives for the spending. "Because Mythos and AI is finding so many vulnerabilities. This is the answer," he said. He acknowledged that rivals may be doing the same, saying "Maybe." He added that Broadcom wants to be "toe-to-toe and better than them." He also drew a line between uses: "As we understand requirements, in some places there's no way out but to use AI; other places just general productivity."

Leadership, not just junior staff

Mahajan is no newcomer to the space. He co-founded Avi Networks, which VMware later bought, and he spent close to a decade at Cisco. When asked whether tooling like this would have helped him when he was starting out, he said the complexity of the work has moved on. Engineers have become compartmentalised, and AI tools may be able to hold more parameters in view at once than a single person can.

"It's difficult for individuals to visualize the whole thing," he said. "So that's where AI really helps. [It] takes all this data and can crunch it together, so it is able to do a better job of that than humans."

He was clear that the cultivation effort isn't limited to new hires. Broadcom wants to encourage engineers at every level, and that includes leaders. "I keep pushing my leaders like, don't be like that. Keep an open mind; otherwise, we'll get left behind compared to our competition," Mahajan said.

What IT teams should take from this

For operations and security teams, the practical lesson is about structure rather than headcount. Pairing a fast, inexpensive first draft with experienced reviewers is a model that other organisations can copy without matching Broadcom's token spend. The open question is whether the review capacity scales as quickly as the AI output does. If Mahajan's numbers hold, signature volume could rise by a large multiple, and someone will have to check all of it.

Watch for two things next: whether the firewall-rule translation agent reaches a product timeline, and whether the 2,000-to-3,500 signature target is met. Neither has a date yet, so for now the useful move is to keep an eye on how review processes change alongside the tooling.

Source: Michael Garcia · www.sdxcentral.com

Same flag