AWS targets hallucinations and AI multi-agents in Bedrock update

AWS CEO Matt Garman overnight announced multi-agent capabilities for Amazon Bedrock – its platform for building and deploying generative AI applications. It’s also taking aim at AI hallucinations and guardrails to ensure safety and accuracy for businesses.

AI Agents are having a bit of a moment in the back half of 2024. September saw Salesforce, Hubspot and Google make big agent announcements within days of each other.

As for AWS, it has flexed the fact it first introduced AI agents in 2023. The latest enhancements take this further by enabling the orchestration of multiple specialised agents within an ecosystem.

Garman emphasised the importance of such tools for increasingly sophisticated workflows.

“While a single agent can be useful, more complex tasks, like performing financial analysis across hundreds or thousands of different variables, may require a large number of agents with their own specialisations,” Garman said during his AWS Re:Invent keynote.

“Creating a system that can coordinate multiple agents, share context across them, and dynamically route tasks requires specialised tools and generative AI expertise that many companies do not have available.”

The new multi-agent feature on Bedrock allows businesses to build specialised agents for specific tasks, as well as an orchestrator agent to manage the overall process. 

For SMEs, these capabilities could present an opportunity to automate complex workflows with minimal development effort. 

For example, a business conducting customer retention analysis could deploy multiple agents: one to parse customer feedback, another to analyse sales trends, and a third to evaluate the effectiveness of marketing campaigns. 

The orchestrator agent could then compile the results into a cohesive report, streamlining the process and saving on manual workload.

AI guardrails and combating hallucinations

AWS has also taken aim at AI hallucinations by building on its previously-existing Bedrock Guardrails feature, which allows businesses to implement safety and responsible AI checks. 

Guardrails now includes ‘automated reasoning checks’, a new safeguard designed to reduce factual inaccuracies, or ‘hallucinations,’ in AI-generated content.

This feature uses mathematical validation to ensure responses align with established facts and customer-specific data.

This technology is particularly crucial in industries like healthcare and finance, where accuracy and audits are critical.

For instance, a health insurance provider could use Guardrails to ensure customer service chatbots accurately respond to inquiries about policy coverage.

The checks can also ensure that model responses adhere to company policies or regulatory requirements.

AWS’ broader push into AI workflows

The announcement was just one of many during the Re:Invent Keynote. AWS also revealed Amazon Nova – a suite of advanced foundation models designed to enhance multimodal capabilities, speed, and cost-efficiency for enterprise customers.

It also announced advancements in its generative AI and cloud infrastructure offerings, including the next generation of Amazon SageMaker and Aurora DSQL.

The author travelled to Re:Invent in Las Vegas as a guest of AWS.

Never miss a story: sign up to SmartCompany’s free daily newsletter and find our best stories on LinkedIn.