Amazon is expanding Bedrock with a preview feature for building conversational agents and several new models from outside the company. The agents are designed to use a business’s own information to respond to requests and carry out tasks, including actions that involve company systems.
Agents connect conversations to business tasks
Bedrock is Amazon’s platform for building generative AI applications with pretrained models from Amazon and startups. Its new Agents feature lets customers configure conversational agents to provide personalized, current answers using proprietary data and to take actions.
One example from AWS machine learning VP Swami Sivasubramanian is a customer service chatbot that processes orders. It could draw on internal information such as customer profiles and return policies to tailor an order to the customer.
That approach links a conversation to information and services beyond the model itself. In a blog post shared with TechCrunch, AWS principal developer advocate Antje Barth described agents as able to understand requests, break complex tasks into steps, ask follow-up questions and make API calls to company systems to fulfill requests.
Instructions shape how the agent works
To decide what information to use and which actions to take, Agents need instructions about their role and task. Amazon’s system generates prompts in the background through configuration steps, including defining what the agent should do and what role it should assume.
For example, a business might configure an agent to help process insurance claims and manage pending paperwork. That role gives the agent direction for handling requests and determining the steps involved.
The feature resembles systems that extend generative AI models with access to external services or data. The source article also points to a broader interest in making AI more useful with enterprise information, including tools from LlamaIndex and Contextual AI.
Bedrock adds models from other companies
Amazon announced additional models for the Bedrock library alongside Agents. Anthropic’s Claude 2 text-generation AI and Stability AI’s Stable Diffusion XL 1.0 image model became available in Bedrock at the time of the announcement.
The platform also added two foundation models from Cohere: Command text generation and Cohere Embed. Command is intended for following instructions and business uses such as summarization, copywriting, dialogue, extraction and question-answering. Embed can support search and classification across about 100 languages.
The combination of Amazon’s own Titan models and models from other providers gives Bedrock customers a wider set of options for generative AI applications. The announcement also ties the platform’s expansion to practical uses that combine a model with company data, tools and workflows.
Bedrock’s reach will depend on access
The larger model selection and the launch of Agents signal Amazon’s commitment to Bedrock and generative AI. Still, the platform had faced questions about customer access. Bloomberg reported in May that six weeks after Amazon demonstrated the technology, most cloud customers still could not use it.
AWS has substantial reach and computing capacity, factors that could help Bedrock find customers. But broader availability and adoption will determine whether the platform’s expanded capabilities translate into regular business use. The announcement establishes what Amazon is offering; the longer-term result remains to be seen.