Amazon Web Services (AWS) this week previewed multiple forthcoming capabilities that promise to make it simpler to manage complex workflows using generative artificial intelligence (AI) agents.
Announced at the AWS re:Invent 2024 conference, these capabilities make it possible for organizations that are using the Amazon Bedrock service to invoke an AI supervisor agent to orchestrate the management of multiple AI agents that need to collaboratively complete tasks in a specific order.
At the same time, AWS previewed an extension to the Amazon Q Business service that promises to make it simpler to build those complex workflows. That capability makes it possible to describe a workflow in natural language that then creates a plan an organization can review, test, modify or approve. Additionally, AWS is providing 50 integrations with third-party applications and services that AI agents will be able to invoke, including Microsoft Teams, PagerDuty Advance, Salesforce, ServiceNow and New Relic.
That integration in the case of New Relic, for example, will make it simpler to apply observability across those workflows, says Camden Swita, the head of AI and ML innovation. While each provider of an application or service will build their own AI agents that will then communicate with an agent to orchestrate a workflow. “There will be AI agents that are specialists for certain tasks,” he says.
AWS has also integrated Amazon Q with its QuickSight service to also make it simple to analyze complex scenarios using a set of natural language prompts.
Finally, AWS is previewing a set of Automated Reasoning checks with the Amazon Bedrock Guardrails service to help you mathematically validate the accuracy of responses generated by large language models (LLMs).
While most organizations at this juncture have no shortage of pilot generative AI initiatives under way, deploying those capabilities at scale has proven to be a challenge, because most business workflows are deterministic in the sense they are supposed to be performed the same way multiple times with 100% accuracy.
AWS CEO Matt Garman told conference attendees that collectively the capabilities AWS is adding will make it more feasible to employ generative AI technologies within business applications. Otherwise, organizations are left to determine what 10% of the output being generated by an AI model might be inaccurate. “It’s been really hard for customers to build inference into their mission-critical applications,” he says.
It’s not clear to what degree organizations are embracing AI agents to automate workflows, but interest is running high. The issue is determining how to review those workflows given the chance there is likely to be some level of inaccuracy that could ultimately negatively impact a customer experience. Even though those errors might be rare, it almost seems inevitable they would be encountered by one or more of an organization’s most valuable customers.
Hopefully, as agentic AI continues to evolve there will be agents capable of validating the workflows created by other agents. Of course, humans performing those tasks may be just as liable to make a mistake, but as a general rule organizations will need to proceed with caution before relying solely on AI agents to automate any specific task.
