OpenAI brings computer control to agents and AWS
OpenAI added computer use to its Agents API at DevDay and joined Amazon to offer Bedrock Managed Agents powered by its models and Codex harness. Developers can now build agents that operate software through the OpenAI service, while AWS customers can create long-running OpenAI agents within their existing cloud environment. The underlying Agents API entered public beta earlier in September; computer control and the AWS route are the latest changes. Amazon describes its joint preview as combining OpenAI models and harness with Bedrock AgentCore, keeping model inference and the agent runtime inside AWS. The distribution agreement reaches enterprises that prefer existing AWS identities, permissions and data infrastructure to a separate agent stack.
The upgraded API brings the Codex harness's multi-agent coordination, tool search, programmatic tool calling and context compaction to third-party applications. Computer use broadens the set of jobs it can complete because an agent can interact with software interfaces rather than depend entirely on a purpose-built API connector. OpenAI runs the underlying infrastructure for its own API service. The developer specifies the model, tools and environment, with choices that include OpenAI-managed sandboxes, the developer's own infrastructure and partner environments. This range matters to a company building agents across legacy applications: it can keep its user-facing workflow while buying much of the orchestration layer as a service.
Amazon's counterpart offering makes a different deployment choice available. Bedrock Managed Agents use AWS identity and defined permissions for each agent, with the runtime and inference operating in AWS and direct access to AWS compute and data services. Amazon calls it a preview and emphasizes continuity of context, memory and specialized work through AgentCore. Salesforce president Joe Inzerillo said joint customers could pair OpenAI models and AWS infrastructure with Salesforce data and controls for production use, adding that ‘data never leaves AWS.’ His comment points to the practical sale: enterprises already control their operations through multiple incumbent platforms, and an agent has to fit those systems before it can earn authority to act.
OpenAI's broader DevDay program also moved Codex into cloud and remote use from any device, with reusable development environments and shared approved settings. These features complement the API offering but address teams using Codex directly, rather than developers embedding the harness in their own product. The original Agents API announcement described no separate fee for the harness during public beta; developers pay for tokens and tools. The new computer-use capability can increase both the breadth and duration of sessions, potentially raising the amount of model work behind each successful assignment. Approval controls and monitoring become part of the product experience when an agent can operate a graphical interface and move across applications.
This launch is best understood as a contest over the agent's control layer. OpenAI wants developers to use its models, its maintained harness and its managed execution, while Amazon offers a route for customers whose cloud commitments and governance already sit in AWS. The joint product broadens OpenAI's reach without forcing a buyer to abandon its cloud operating model. It also gives Amazon a frontier-model agent offer under its own account and permission system. Developers can compare both channels against a self-built harness, a rival model provider or another cloud platform. The choice will turn on reliability, security reviews, integration effort and the total cost of a completed workflow.
Analysis
OpenAI is selling more of the agent stack than inference alone: a maintained harness, context management and tool execution can reduce customers' engineering burden and make its models harder to swap out. AWS can supply distribution and procurement credibility, but it owns the cloud account and much of the customer relationship in Bedrock Managed Agents, giving it bargaining power over economics and positioning. The OpenAI API route offers a closer developer relationship; the AWS route can reach buyers unwilling to relocate data or governance. Computer use increases addressable workflows and potentially token consumption, while the price of operational mistakes rises with the scope of an agent's permissions.