If you go on LinkedIn, you’ll see hundreds of AI influencers saying they were able to get AI agents to do complicated workflows across multiple pieces of software. So why could we not get Operator to do something as simple as order a bag of chips?

The reality is, a lot of people are using the term ‘agent’ when what they really have are logic-driven workflows. We sat down with Jiquan Ngiam, co-founder and CEO of Lutra AI – a platform enabling users to build AI agents – to level set on what’s really possible with agents right now.

What is actually true about AI agents today

Today’s agents are capable – but stunted by the world they were developed in. Jiquan says they excel at language and data: processing it, understanding it, working with it, and figuring out how to get it into a usable format.

These capabilities work best with very prescribed workflows: “Go into my call transcripts, pull this client’s information, put it into Salesforce.”

“That's working to a point of very high reliability now,” Jiquan says.

Where agents fail (for now) is use cases that require them to work autonomously on your computer. The idea of AI being able to move across different softwares and accomplish whole tasks is still a pipe dream – and that’s because the tech isn’t designed to do it yet.

“In the systems we have today in the market, the APIs are designed for two audiences: Users and developers,” Jiquan says. “The challenge is figuring out how to take the current interface and design it to allow AI to connect to it, too.”

This is the emerging opportunity of AI agents, and companies like Lutra AI in the application layer are slowly chipping away at it. But the reality: This advancement is not exactly around the corner.

The 3 maturity levels of AI agents

Jiquan broke down the AI hype into three levels of agentic maturity based on reliability – in other words, how often you can get the AI agent to do something correctly.

Level 1: Existing capabilities with high reliability

This is the set of agentic capabilities that are working today.

• Highly prescribed workflows – e.g. pull contact info from a website → update a Google Sheet.
• No autonomous decision making – executing narrowly defined workflows that follow a set, repeatable structure.
• Focuses on an LLM’s core strengths – transforming data across formats and channels.

Level 2: Emerging capabilities with moderate reliability

Agent capabilities that are currently emerging, are limited by today’s interfaces, but are feasible when considered in the development of new tools.

• Access to any tool – agents that can access our everyday software without limitation.
• Coordination across software – agents that can connect systems and exchange data between them.
• Higher level access to tools – Instead of just clicking buttons or calling low-level APIs, agents that can perform higher-level commands like “create spreadsheet”.

Level 3: Agent hype with low reliability

Agent hype that is still completely experimental due to highly error-prone and unpredictable tech.

• Fully autonomous AI – agents that control a full desktop environment or browse the web independently.
• AI coworkers – agents that make decisions, click around UIs, and respond dynamically to their environments.
• Complete reliability – agents that can perform multi-step, adaptive tasks without messing up somewhere along the chain.

“A lot of those things are just not reliable enough yet,” says Jiquan. “But they’re very fun to look at and very exciting to see the promise of. We’ll get there eventually but not any time soon.”

The current opportunity in AI agents

AI is unpredictable because it’s probabilistic by nature. You can give ChatGPT the same prompt 60 seconds apart and get different answers. So one of the main challenges in building agents is designing them to be able to do the same things over and over.

“If your car only works five out of seven days and doesn’t drive on weekends, you’re not going to buy that car,” Jiquan says.

One work-around is a long series of detailed prompts that  tries to account for every situation the agent might encounter – but that’s essentially impossible.

“There’s no way to guarantee an AI will follow an entire long prompt correctly each time,” says Jiquan.

The other work-around is to leverage the strengths AI already has – including writing code.

Lutra’s approach to making AI agents more deterministic is to have the AI write a program that executes on the actions the user prompted, and then run that program. So you say “comb Linkedin for anyone with this title and put their name, company, and profile URL into a spreadsheet” and the AI generates code that accomplishes that task.

Users have access to the code the AI has written and can ask it to make adjustments – then the agent runs the same code every time with no deviations.

The hype-free view of the future of agents

We asked Jiquan for his honest predictions on the future of agents, as someone with boots on the ground in the AI agent space. Here’s what he thinks you should expect:

• Predictability > autonomy. Everyone wants the flashy agents that can do it all, in theory. In practice, reliable, deterministic systems will be way more desirable. So expect agents that can’t do everything, but can do a few things really well.
• We’ll all be managers to AI. Agents will still require supervising, tweaking, and validating. That will be our new role in the tasks they take over.
• We’re years away from full autonomy. Jiquan calls these agents ‘research-grade’, because they’re mostly theoretical. So AI coworkers are more a concept than anything else at this point.
• Agents = integration hubs. The power of agents is in connecting them across systems. Jiquan sees a world in which this intelligence is part of our computer helping to tie it all together for us.