Mere days after some of the most prominent ‘big AI’ advocates went on record to warn of an existential threat to humanity arising from unchecked AI development, Splunk leadership made improving trust in agentic AI the headline topic at its annual .Conf customer event in Denver, Colorado.
Taking to the keynote stage, Jeetu Patel, chief product officer and president at Cisco, set the ball rolling by reflecting on the momentous change that has dominated the technology industry since the dawn of the agentic era in the autumn of 2024.
“We have squarely moved now from the era of chatbots to the world of agents, conducting tasks for us almost fully autonomously,” said Patel, who urged the audience to think of agents not as tools but as teammates.
Patel described four key changes underpinning the rapid growth of agentic AI. Firstly, he said, “inference is the new workload”. Chatbot infrastructure consumption was spiky and ultimately, gated by a human’s capacity to engage – even power users weren’t putting too much strain on the underlying infrastructure, but now, agents are working 24-7, spawning more workloads for other agents almost continuously, so demand patterns are now higher, and more consistent.
“In February this year the tokens consumed by agents for the first time exceeded the total consumed by humans, which is wild,” said Patel.
Seven short months later agents consume five times as many tokens, and since very few humans could currently be described as power users, this will only increase.
Secondly, he continued, “agents are the new workforce”. This year, claimed Patel, 60% of global AI compute capacity will be put towards compute, with billions – maybe trillions – of agents using over five times more network bandwidth than humans, and maintaining visibility into the network to control their behaviour is becoming critical.
“Agents are like teenagers,” said Patel – they are intelligent, fearless, and have poor judgment. They need to be monitored, and any drift in behaviour needs to be corrected in real time. That requires a fair amount of emphasis on visibility, as Patel observed, “you can’t fix what you can’t see”.
Thirdly, he said, “tokens are the new currency”. The cost of tokens in the agentic world is in danger of getting out of control as fast-moving agents get chatty with one another in a way that humans talking to chatbots could never mentally sustain.
As a result, said Patel, we should anticipate more emphasis on open source and open weight AI models, and greater reliance on inferencing at the network edge, not in the hyperscale datacentre. In fact, predicted Patel, those building out inferencing infrastructure will start to embrace a hybrid model, and while hyperscalers will remain a strong area of growth, organisations will build capacity in on-prem datacentres and even desktop computers.
Finally, “control is the new moat,” said Patel. We now live in a “post-Mythos world” where safety and security have become ironclad necessities, and all of the assumptions people had previously made about AI security are changing.
The bottom line, he said, is that every agentic action is simultaneously a routing challenge, a trust decision, and a telemetry event. This is specifically why, Patel claimed, Splunk and Cisco created a vertically integrated, full stack platform, from silicon and photonics, up to network, compute, data, models, and AI apps and agents, with observability and security present at every level so that when the worst happens – whether that’s a technical outage or a cyber breach – detection and investigation becomes child’s play, and resources can be directed to remediation.
Ultimately, said Patel, this means organisations will be able to run agentic AI in confidence. He described a “world class, best of breed” observability proposition giving visibility into agent activity, application runtime, model performance and infrastructure utilisation; an agentic security operations centre (SOC) with the ability to process and understand every security event; and telemetry correlated from multiple domains, federated and analysed at lower cost, and unlocked for model training and agentic action.
This comes together in Cisco Cloud Control, which was unveiled at Cisco Live back in June and became available to US customers at the end of August. Patel described it as centralised control plane based on an agent-first operational model.
“Delivering digital resilience starts with Splunk. This is one of the most strategic acquisitions we [Cisco] ever made. If we were to do it again I probably could be tricked into paying more for it,” said Patel.
