Cisco's Automation Chief on AI Agent Guardrails and the 100x Engineer Myth
- Joao-Pierre S. Ruth
- 525 views
Cisco's chief automation officer says coding and testing are speeding up, but design and deployment decisions still slow teams. His guardrails for employee AI agents are worth a look.

Srini Namineni, Cisco's chief automation officer, has a view on AI agents that IT leaders should weigh before they hand out access. Give staff vetted, tested agents and productivity rises. Let those agents touch the wrong systems and you've built a new problem. His position is that both things are true at once, and that the gap between them is where most of the work sits.
Code is fast. Deciding is still slow.
Namineni has been with Cisco for eight years and became chief automation officer in September 2025. He leads IT automation strategy and the company's internal AI work, including autonomous agent deployment across staff. He describes generative AI as a "generational technology," and says bringing it into Cisco's own operations is part of the job he was given.
He pushes back on the idea that AI turns every engineer into a 10x or 100x performer, the sort of rockstar developer who produces exponentially more than a typical peer. The more useful question, in his framing, is whether production and the market can absorb that speed. Companies want to clear the bottlenecks that hold them back, and AI could open doors to new markets, new products and different sales channels. But only if the rest of the machine can keep up.
His summary of the bottleneck is blunt: "The code piece is getting faster. The testing piece is getting faster. But the thinking of what to design, how to deploy — those are still blockers."
That's the part I'd underline for any IT manager planning around AI coding tools. Faster output at the code and test stages doesn't automatically shorten the release cycle if architecture reviews, deployment decisions and change approvals still run at human speed. Teams in this position often find that the queue simply moves downstream.
The backlog problem gets harder, not easier
Namineni expects businesses to spend time catching up. Their infrastructure and operations need updating to support engineers working with AI, and tech debt and vulnerabilities will need faster attention as output rises. More code moving through the pipeline means more code that needs to be secured, patched and maintained.
His view of the starting point is hard to argue with. "You ask any IT department, nobody says, 'I don't have backlog.' They have years of backlog," he said. If AI speeds up new work but leaves that backlog untouched, the organisation ends up with more change to manage on top of the old debt. My read is that the teams who benefit most will be the ones that use AI on the backlog itself, not only on new features.
What Cisco actually gave employees
In late August, according to Namineni, Cisco made an in-house personal AI agent available to every employee after a period of testing and evaluation. The agent is designed to act only at the employee's direction. In practice, the typical jobs are fairly mundane: reviewing calendars so staff know what to prepare for, finding a conference room for a meeting, or booking vacation time.
Namineni was specific about where the human stays in control. "It can do approvals, but the human is in the loop," he said. "We are making sure that the humans are working with them and delegating tasks that are either low risk or completely approved by them."
That's a sensible pattern to copy, and it's easy to state. The harder part is defining "low risk" in a way that holds up across departments. A conference room booking and a purchase approval both look like "tasks," but they carry very different consequences. Any rollout should start by writing down which action classes an agent may perform without a person signing off, and which ones it may never perform.
Costs, rogue actions and the payment line
Namineni also raised money and misbehaviour. He warned about escalating costs as agents consume tokens, with bills that can arrive unexpectedly. He also pointed to rogue actions by agents as a live concern. He doesn't think either risk is a reason to sit out the technology. "We have taken the leap of faith, and we are constantly checking in the background," he said.
The background checking is the operational piece. Token consumption is a metered cost that can grow quietly, so usage visibility is a practical requirement rather than a finance nicety. Agent actions need logging and review. Nobody should assume a tool stays within its brief because it did so last week.
Namineni's technical explanation for why agents go off-script is worth understanding, even if you aren't building models yourself. He says the underlying technology is probabilistic, and that the creativity people see in generative AI and large language models comes from that same property. When an agent appears to act outside its instructions, he links it to the model trying to reach its goal by whatever route it finds. That's why limiting access matters so much. His advice on that point was short: "The key is you don't connect these agents to payment systems."
I'd treat that as a starting line for access design. Agents should only reach the systems their specific tasks need, and money-moving systems sit outside that set entirely unless there's a very deliberate and reviewed reason to change it.
Thinking is the skill to protect
The last theme is the one that's hardest to measure. Namineni warned against letting skills and determination wither and simply accepting whatever AI proposes. AI answers arrive quickly and sound confident, which makes it easy to skip the work of reasoning through a problem. He described that slide as a slippery slope.
"Coding is OK in my mind, but not like, 'I don't know what to wear this morning. Tell me.' We don't want to go to that point," he said. And then the line worth pinning above a team's desk: "Coding is a skill that you can outsource, but thinking in general as a skill we should never let go."
For managers, that raises a practical question about review and training. If junior engineers accept generated code without understanding it, the organisation may find out only when something breaks in production. Code review, design discussions and incident postmortems become more important as AI output rises, not less.
What to check before the next rollout
Namineni's comments point to a few questions any IT team can ask this quarter. Which agent actions are pre-approved, and which need a human sign-off? Which systems can an agent reach, and is any payment or financial system within that reach? Who sees token spend, and how often? Are agent actions logged in a way that someone reviews? And is the backlog being addressed alongside the new work?
Cisco's approach is one large company's experience, described in its own words. It's not a template. But the sequence he describes, starting with vetted tools, keeping humans in the approval loop, watching costs and access, and protecting engineering judgement, is a sound checklist for teams that are still deciding how far to let agents run.