AI-native agents turn to model-driven design – News

As organizations look to roll out artificial intelligence initiatives, they’re quickly discovering that the fixed workflows and step-by-step procedures that work so well in traditional programming come up short when designing models for AI-native agents. 

Because agents operate in open-ended, dynamic environments and make decisions the developer did not explicitly script, trying to specify every step in advance is counterproductive. A better idea is to work with the model, take advantage of its capacity for reasoning and allow it to generate the needed logic dynamically, according to Clare Liguori (pictured), senior principal engineer at Amazon Web Services Inc.

“For complex tasks in the past — when models weren’t as capable — people would go back to those same programming paradigms that they’re familiar with [and] create a workflow. As it turns out, that’s a very brittle approach,” Liguori told theCUBE. “When [AWS] went back to the model-driven approach, we found that we can actually guide and steer the agent and let the model come up with the workflow itself, and still have the capability to do really complex tasks in the agent. It’s a very different mindset among younger companies, younger developers, that are coming up in this completely non-deterministic AI world.”

Liguori spoke with John Furrier at AWS re:Invent, during an exclusive broadcast on theCUBE, News Media’s livestreaming studio. They discussed how the need to simplify model development for AI-native agents is impacting traditional programming techniques.

Overcoming structural challenges with AI-native agents

A major challenge in the development of AI-native agents is the large amount of “structural glue” — such as orchestration code and guardrail logic — that goes into model development. Instead of building the model, developers spend most of their time on the boilerplate surrounding it, which can account for as much as 90% of enterprise AI efforts. This is not just a structural inefficiency that increases the cost of AI engineering, but it actually leads to serious quality issues, according to Liguori. 

“We found that frontier models, especially, are so capable of reasoning and driving the tool selection and things like that in the agents, that the more that you do around it, the worse your agent gets,” she said.

Not having to focus on the structural  “surround” also simplifies ongoing agent management, she added. This shift lets teams focus on business logic and outcomes instead of constantly reworking low-level plumbing.

“As the models are getting better, which we’re seeing every few months, you don’t have to do anything to your agent, other than change which model you’re using for your agent to suddenly get better,” Liguori explained. “You don’t have to change your entire software stack around it.”

The recognition that the inherent complexity involved in model development can be an obstacle to enterprise deployment of AI-native agents was a big factor behind AWS’s decision to integrate TypeScript as an orchestration language for building agents, using its open-source AI agent framework Strands. Taking advantage of TypeScript opens the door for model development to a much broader set of developers, Liguori noted.

“We wanted to make it so easy for anybody who is not an AI expert to be able to write an agent in a few lines of code,” she said. “I’ve had product managers come to me and say, ’I wrote a Strands agent, and this is amazing.’”

Here’s the complete video interview, part of News’s and theCUBE’s coverage of AWS re:Invent: 

https://www.youtube.com/watch?v=JhJPMGkqJjQ

Photo: News