At some point during the 2010s, a meme surfaced wherein people would see someone on a skateboard, point a camera at them, and call out, “Do a kickflip!” It was a kinder call from the car window than the “skate or die” that dominated the ‘90s, but DAK is pressurized. I’ve only landed a handful of kickflips in my lifetime, and if a stranger implores me to do one, part of me wants to perform.
We’re in a similar moment right now with technology, as organizations and individuals are being asked to do something dazzling with AI. “Build an agent” has become the equivalent of “do a kickflip” — often well-intended, but usually blind to the complexity behind completing the task.
It can take months (or years) of trying for a skateboarder to land a single kickflip, but that doesn’t mean that each attempt that doesn’t land is a failure. They are often learning opportunities — Wha? I relaxed my ankle that time and was able to flick the board better!
It might also take hundreds of attempts to get AI agents working effectively, but every agent that doesn’t behave as intended can get you closer to building one that does. It’s making the API call at the right time, but I need to check the JSON in the MCP and try again. When teams are building with frameworks and runtimes and can crank out new iterations in a matter of hours, or even minutes, learnings can be poured directly into new versions. In this way, the organizations making the most progress with agentic AI are trying kickflips all day long.
Failing forward is failing
Agentic AI relies on a fast interaction process that’s often described as being “more agile than Agile.” Beware when skating down the proverbial sidewalk. Common adages around “failing forward” and “moving fast and breaking things” are the invisible pebbles that lock up your wheels and send you flying face-first into the pavement.
With agentic AI, the immediate objective isn’t to build an AI agent that can run an entire business. The first step is figuring out what to automate and what not to automate. This can be a difficult thing to gain clarity on, and there’s an understandable urge to use AI agents to automate existing processes. Using AI agents to automate processes that are already being handled by robotic process automation (RPA) might offer some meager improvements by introducing a conversational interface, but it often means inviting complexity for complexity’s sake, without providing a whole lot of value — failing forward.
McKinsey research suggests that 90% of pilots fail to reach full production due to the complexity of the task at hand. “Enterprises need to get many elements right simultaneously — address unstructured data, develop advanced algorithms, build the right IT architecture, drive capability building, change management, and domain expertise, to name a few,” says senior partner Ben Ellencweig. “An ecosystem is the best way to stay current and win in this space; it moves too quickly for prior conventional methods.”
So while the sentiment of “fail forward” is correct — you do have to be willing to take chances — it’s dangerous to adopt the mindset that by simply doing anything at all with AI, you will be making progress. Organizations taking agentic AI seriously are looking for agent runtimes that can:
- Maintain agent memory and goals across interactions.
- Use APIs, databases, and webhooks to enable access to external tools.
- Facilitate multi-agent orchestration.
- Handle input and output across text, voice, UI, sensors, etc.
- Operate continuously in the background.
An agent runtime gives organizations a place to experiment freely and create new experiences using agentic tools. It allows teams to have ideas, prop up agents, and start tinkering and testing right away. It’s what allows them to become more agile than Agile.
Don’t be a kook
For skaters, a kook is someone posturing like they know how to kickflip when they can’t even stand on a board. Right now, most organizations are in full-on kook mode when it comes to AI, using bolt-on point solutions to make it look as though they are pursuing something real. Many of the platforms and solutions purporting to sell agency in the form of AI agents are also poseurs. They’ve spent the majority of their time and money on looking the part without investing in the mechanics of riding the thing.
Gartner has estimated that 40% of agentic AI projects will fail in a few short years due to high costs and low ROI. As Gartner sees it, agentic AI is more hype than substance. They say that using agentic AI without strategic alignment, business relevance, and budget-conscious deployment will lead to more failure, and that 2027 will be the year of reckoning. In other words, most deployments will fall victim to kooks.
- Kooks will use the wrong success criteria … I need to do a kickflip first-try.
- Kooks will try to automate the wrong thing … Maybe I’ll try kickflip on a bicycle.
- Kooks will add unnecessary complexity … I should blindfolded double kickflip first.
- Kooks will see AI agents as isolated entities, not members of an ecosystem … Can I kickflip with nothing but four urethane wheels?
Companies that can’t break through the siloes created by traditional software will end up paying a bunch of money to look foolish. Meanwhile, there are people inside most organizations already experimenting with agentic AI. Many of them have good ideas about how to improve existing workflows in meaningful ways.
When these people are actively involved in the creation, testing, and adoption of AI agents, orgs stand a much better chance of building and evolving successful agentic ecosystems. But this only happens when team members can see the value of automation and are empowered to contribute to the creation of agentic solutions. These people have taught themselves how to kickflip and can show business leaders how not to be kooks. In turn, business leaders can give them more skateboards and an adequate environment to push the limits on how they can use them.
Find balance
On their own, AI agents are sloppy generalists, powered by elusive and often unreliable large language models. Much like kooks, LLMs are good at pretending to know what they are talking about, which is a huge problem when it comes to safe and reliable adoption. Therefore, one of the things an agentic ecosystem needs in order to function properly is a knowledge base with access to verified information about its home organization. Otherwise, AI agents will just make shit up.
Back in the pre-ChatGPT world, JP Morgan spent nearly a year creating an agent for their advisors, working directly with Sam Altman and OpenAI. Their team audited 60,000 internal documents and built an accurate knowledge base so that their agent could provide advisors with accurate, relevant information in real-time. This required systems for keeping the information updated using a time-to-live approach, where documents come up for review and must either be updated or removed from the knowledge base.
Automating complex workflows in truly novel ways is a whole lot more than pushing around on a board in the driveway. Most organizations will require an agent runtime environment that offers code-free building tools that make it easy to automate conversations with employees and customers. As in traditional software, a runtime is where teams can execute AI agents, testing them and iterating on them, and creating increasingly sophisticated orchestrations to automate real work.
In skateboarding terms, an agent runtime is like a big warehouse with piles of wood and power tools in one corner. If people have ideas for things they like to kickflip over, off of, or into, they can build them and test them out. The things that work well can be sequenced along with other creations, letting orgs kickflip off a kicker ramp and into a salad grind, so to speak. Best of all, when you’re ready, you can open parts of it up to the world, while continuing to experiment and innovate.
Do a kickflip
Ultimately, you want to take your kickflip out into the world and test it over higher obstacles and wider gaps. What gets overwhelming about agentic AI is that there are objectives nested inside objectives. There’s the objective of learning to ollie, the objective of learning to flick the board with the edge of your foot, the objective of lifting your back leg up and out of the way, then the objective of landing the kickflip, then the lofty objective of giving it style and authority.
Also consider that the aforementioned “skate or die” era is back upon us. Whether they like it or not, orgs are in the position of getting busy kickflipping or getting busy dying. Gartner has recently predicted that by 2029, AI agents will resolve 80% of common customer service issues without human intervention. They say that this will open up a 30% reduction in operational costs, which amounts to a massive advantage.
Every pathway to success will look and feel different, and they will only emerge as teams begin to explore the limitations and powers of agentic AI within their unique businesses and industries. It will be messy, weird, and embarrassing at times, but by establishing a clear vision for how you want to use agentic AI, getting serious about finding partners who can help you use these tools holistically, and creating a proper environment in which AI agents can do their thing, bold-ass companies will do the kickflip first and skate right past their competitors.
Featured image is AI-generated.
