Cast your minds back to January 2025. Darts prodigy Luke Littler became the youngest-ever world champion at just 17 years old. Donald Trump was gearing up to be sworn back into office for his second term. And South Korea's suspended President Yoon Suk Yeol was finally arrested following the previous year’s martial law crisis.

But that same month, OpenAI co-founder Greg Brockman took to X (formerly Twitter) and made a prediction: 2025 will be a year defined by a shift away from AI being chatbots to agents.

A daring act of divination from the former Stripe CTO, but it was a line Brockman and other OpenAI bigwigs would repeat ad nauseam throughout the early months of the year, culminating in the reveal of Operator, a browser-based agentic AI system built on one of its now-lesser-known AI models: Computer-Using Agent (CUA).

At that point, the concept of AI agents was still rather new. OpenAI had been working on some related projects for some time, rolling out early agentic features in its ChatGPT app back in May 2024. That same month, Google DeepMind showcased Project Astra, teasing their efforts to build a universal AI agent tool capable of performing a multitude of tasks.

But it didn’t take long for Brockman’s comments to take hold, and the industry ran with the hot new thing. Move over, foundation or frontier models. Out of the way, domain-specific AI. Mind your head, small language models (SLMs) – there was a new "next big thing" in town.

Vendors everywhere raced to put out with agentic solutions. In the world of networking alone, there was Gluware Titan, Cisco's AI Canvas, the Network Topology Agent from Articul8, and Google’s CogniPort. And it didn’t stop there. Netcracker, Nokia, IBM, and Hewlett Packard Enterprise (HPE)/Juniper all came out showcasing some form of agentic offering.

So, Brock-stradamus, right? Was 2025 really the year of the AI Agent? And what’s in store for 2026?

But first, what is agentic AI?

Agentic AI systems are defined by their ability to pursue a specific objective by autonomously making decisions and initiating actions. Rather than waiting for a command or running a fixed script, such systems are capable of assessing what’s happening, deciding what needs doing, and getting on with it.

In a networking scenario, this could mean a tool capable of spotting system faults, like sudden spikes in network latency, or isolating devices exhibiting unauthorized behaviors, and applying a fix without the need for human approval or assistance.

Unlike the ChatGPT of yore, an agentic system would ingest vast quantities of hyper-relevant data to help it perform and execute assigned tasks. And compared to intent-based networking, which focuses on enforcing predefined policies, an agentic system should be able to interpret intent dynamically, reasoning about context, trade-offs, and goals as conditions change.

Agentic AI vs automation, intent vs instruction

Explanation aside, then, was 2025 truly a breakthrough year for AI agents?

SnapLogic CTO Jeremiah Stone said 2025 was the year when creating scalable, production-ready agentic systems took off.

“Context engineering, creating structured, repeatable short-form specs that serve as the shared automation language between humans and models, was the hot skill for AI developers,” he added.

Akhil Verghese, the founder and CEO of AI consulting agency Krazimo, looked not to Greg Brockman, but another OpenAI co-founder: Andrej Karpathy, who said the next 10 years will be the decade of agents.

“I believe 40-70% of all white-collar work will be automatable within three years,” Verghese told SDxCentral. “But there's a big gap between ‘automatable’ and ‘automated.’ I think the 10-year timeline is more realistic (maybe an overcorrection) to see nearly all white-collar work as it is today completed automatically. I feel certain there will be new things to do, but those are more difficult to imagine.

“2025 was more the testing and experimental phase, and I think 2026 is going to have some of that too. It was also the year of companies learning painful lessons about adopting AI solutions without adequate guardrails, success criteria, and maintenance plans.”

Some early use cases that emerged included Cisco demoing how security-optimized AI models could be tied to an agentic framework to power an autonomous agent for security operation center (SOC) applications. Another example saw HPE sprinkle agentic AI management capabilities into its GreenLake platform in a bid to boost better coordinated data flows between a user’s preferred hybrid IT environment and their AI agents.

Jean-Philippe Avelange, CIO at managed network-as-a-service (MNaaS) provider Expereo, agreed with the consensus that 2025 was more about experimenting, with agent concepts layered on top of fragile foundations, including “fragmented data, brittle workflows, and infrastructure that was never designed to be autonomous.”

“That experimentation was necessary, but it exposed a deeper issue that extends beyond technology: most organizations cannot clearly define what 'correct' means,” Avelange added.

The rise of ‘agent washing’

In spite of the preposterous pace of proliferation of agents, a key report also emerged in 2025 that seemed to suggest there was more to agentic AI than meets the eye.

Published in June, Gartner’s report projected that more than 40% of agentic AI projects would find themselves canceled by the end of 2027. The analyst firm suggested that by 2028, just 33% of enterprise software applications would feature some form of agentic AI. That’s up from less than 1% in 2024, though it is hardly a meteoric rise.

But the real kicker from that report was the realization of “agent washing” – where firms simply take existing offerings like AI assistants, chatbots, or robotic process automation (RPA) tools and slap the agentic label on them. Of the thousands of agentic AI vendors out there, Gartner’s report estimated that really only 130 of them were the real deal.

Expereo’s Avelange told SDxCentral that the idea of agent-washing was “an inevitable consequence of hype outpacing operational readiness,” one seen previously with automation offerings, because the market rewarded the label rather than the outcome.”

“The next phase will be less forgiving,” Avelange said. “Buyers are now asking whether agentic systems reduce cycle time, improve resilience, or lower operational cost in production environments. If the answer is no, the terminology no longer matters.”

From a network and infrastructure standpoint, Avelange argued that agentic AI will persist only where it is embedded in platforms designed for autonomy, observability, and recovery – while being quietly reverted to assisted workflows elsewhere.

“In networking, genuine agentic AI is not a chatbot bolted onto a dashboard,” the Expereo exec said. “It is software that can reason over constraints, negotiate trade-offs, trigger actions across suppliers, and continuously optimize outcomes based on real-world conditions. That requires structured data, programmable workflows, and a network fabric that is observable and controllable end-to-end.

“In 2025, we observed widespread agent-washing: existing RPA scripts and assistants were relabelled as ‘agents’ without true autonomy, decision boundaries, or accountability," Avelange continued. "If a system cannot safely act, adapt, and recover without constant human babysitting, it is not agentic. It is simply automation with better marketing.”

Caleb Moore, co-founder and CTO of the AI fraud prevention firm Darwinium, argued that most agentic AI projects “have been sincere, rather than agent washing,” but that the development of genuinely good agentic tools has been far harder than most developers would have assumed.

“Interacting with a large language model (LLM) through context and tools is a fundamentally different problem space than anyone had experienced, including AI experts, who are familiar with the mathematical foundations but equally unprepared for the emergent behaviors they have brought forth,” Moore said.

“This has been exacerbated by rushing products to market to catch the wave, leaving many offerings unable to deliver on their own visions. The dominance of top-down mandates driving AI transformation throughout 2025 has made technology benchmarks and check-box feature list[s] the key focus on both the supply and demand sides, rather than practical utility.”

With every new technology cycle, of course, there’s always a metric ton of hype. Going into 2026, Kaushik Shanadi, co-founder and CTO of end-to-end security platform Helmet Security, suggests the new year will see agentic adoption “speed up.”

“Gartner’s warning is cautious, as there will be a lot of new players and consolidation is normal in emerging technology," Shanadi said ”However, the difference this time is that end user experience will be the clear separator.”

So what does 2026 have in store?

If 2025 was the year of experimentation and a fair bit of overpromising, then 2026 may just be when the rubber meets the road.

Genesys CTO Glenn Nethercutt told SDxCentral that 2026 will be "the year AI stops observing and starts operating."

He pointed to a shift from LLMs to large-action models (LAMs) – systems that don't just chat, but actually do things.

“Expression without execution is reaching its limit. The next generation of systems will move with purpose, transforming understanding into action through LAMs,” Nethercutt said. “This signals the rise of agentic intelligence – AI that engages with the world instead of merely representing it.

“Powered by LAMs ... the next generation of AI will change state before friction emerges: routing, advising, escalating, and intervening based on intent signals, microcontext, and temporal patterns instead of after-the-fact defects," Nethercutt added. "Journeys will be shaped in advance rather than repaired retroactively. “

But it's not just about flashy new model architectures. Nikolaos Vasiloglou, VP of machine learning research at RelationalAI, argued that the real game-changer will be agentic memory: systems that can actually learn and build on past interactions rather than starting fresh with every prompt.

Avelange, meanwhile, predicts 2026 will be less about deploying more agents and more about being smarter with the ones we've got.

“It will be about delegating fewer responsibilities to them and clearly defining their scope,” he said. “That trust will only exist where organizations can express intent precisely, define correctness upfront, and close the loop between decision-making, execution, and recovery.”

Avelange also provided his own prediction for the infrastructure space: agentic development tools could drive wider adoption of Infrastructure-as-code (IaC) by making advanced features of tools like Terraform more accessible to network and infrastructure engineers who don't fully embrace coding.

"This can allow for rapid evolution while maintaining deterministic rollout and rollback," Avelange added.

Brockman’s prediction then was ultimately right, sort of. Last year was indeed when the industry got serious about agents – or perhaps jumped on the bandwagon. But the last 12 months were also about firms testing their development limits and learning some hard lessons about the gap between autonomous and actually useful.

This year, then, looks set to be when we find out which of those lessons stick and which vendors were serious all along.