OpenAI and AWS unify agent workspaces

OpenAI and Amazon Web Services spent the same week pushing AI agents out of one-person chat windows and into shared company workflows. (openai.com) (aws.amazon.com) On April 22, OpenAI introduced workspace agents in ChatGPT, describing them as Codex-powered agents that teams can share, run in the cloud, and connect to business tools. OpenAI said the feature is an evolution of GPTs, with organizational permissions and controls applied to what agents can access and do. (openai.com) OpenAI’s business page says those agents can automate multi-step work across Slack, Google Drive, Microsoft SharePoint and other enterprise systems, with governance and admin controls layered on top. Third-party coverage said the rollout also reaches Salesforce, Notion and Atlassian Rovo integrations for business users. (openai.com) (venturebeat.com) Amazon Web Services made a parallel move in Amazon Bedrock AgentCore. In an April 22 post, AWS said its new managed agent harness lets developers define an agent’s model, tools and instructions, then run it in three API calls without writing orchestration code. (aws.amazon.com) AWS paired that with a new command-line interface, prebuilt coding skills and a persistent filesystem so agents can pause and resume long tasks. AWS said the managed harness is in preview in four regions, while the command-line interface and filesystem are available globally. (aws.amazon.com) The technical problem both companies are tackling is the same: an agent that can act for multiple employees needs memory, tool access and guardrails that survive beyond a single chat session. OpenAI is packaging that as a shared workspace inside ChatGPT, while AWS is packaging it as infrastructure developers can wire into their own apps. (openai.com) (aws.amazon.com) Security is the point where the approaches meet most clearly. AWS documentation says AgentCore Runtime isolates each user session so credentials and permissions do not spill across sessions, and AWS describes the runtime as framework-agnostic and model-agnostic. (docs.aws.amazon.com) (github.com) OpenAI is making a similar pitch in product terms rather than infrastructure terms. Its launch post says workspace agents operate within organizational permissions, and its business page frames the product around governance, admin controls and secure connections to company systems. (openai.com 1) (openai.com 2) The contrast is also the market map. OpenAI is selling a finished workspace inside ChatGPT for teams already living in its software, while AWS is selling the plumbing for companies that want to build and host their own agents on Bedrock. (openai.com) (aws.amazon.com) That leaves enterprises with a clearer split than they had a week ago: buy the agent workspace, or build on the agent runtime. In both cases, the product being sold is no longer just a chatbot, but a managed place where software can do work on behalf of a team. (openai.com) (aws.amazon.com)