AWS Takes AgentCore to General Availability and Bets Big on Autonomous Agents

In New York, on June 17th, Amazon Web Services announced the general availability of the AgentCore harness, the orchestration engine that runs agents in Firecracker microVMs, isolating each session in a dedicated sandbox. The AWS Vice Presidency of agentic AI, led by Swami Sivasubramanian, delivered the keynote at the Javits Center in front of over 200 technical sessions. Along with the harness, AWS also released Web Search in AgentCore and Policy with Bedrock Guardrails in GA, two items that were lacking for enterprise customers to run agents in production without human firewalls in between.

Southwest Airlines emerged as the standout case study, detailing the use of agents in real operations to support crews and airport management. It is rare for a company of this size to detail agentic AI in production outside of marketing slides, and this segment served as the only grounded reference in an event marked by demos.

Quick, Transform, and the Assault on the Desktop

On a parallel front, AWS launched autonomous capabilities in Amazon Quick, the assistant that learns user workflows within corporate applications. The pitch is for the user to describe a task in natural language and adjust the level of autonomy, from step-by-step approval to goal-based execution. As an example, AWS showcased a financial agent processing orders as they arrive and a sales agent monitoring CRM, emails, and Slack to signal risk in accounts and suggest next steps.

However, the most ambitious piece is AWS Transform with continuous modernization, still in preview. The service scans code repositories against configurable baselines, generates findings within hours, and opens autonomous pull requests to remediate obsolete dependencies, old frameworks, and vulnerabilities. According to AWS, the move to the CI/CD pipeline can reduce operational maintenance costs by up to 30%. In a year of operation, AWS claims that Transform has processed 4.5 billion lines of code and saved 1.6 million hours for customers of the previous program.

The Competition with Microsoft, Google, and Anthropic

For AWS, the GA of AgentCore is a defensive move. Microsoft has moved Copilot Studio forward as the backbone of corporate agents integrated into Office, and Google has pushed Vertex AI Agent Builder. Anthropic, now at $30 billion in annualized revenue and a key model within Bedrock, drains much of the mindshare in enterprise agents. Without a GA harness, AWS risked becoming cheap plumbing with no voice at the product level.

The strategic play here is Bedrock as a neutral gateway. On June 1st, AWS brought GPT-5.5 and Codex to Bedrock in GA, alongside Anthropic models. For the CIO, this means running Anthropic, OpenAI, or proprietary models under the same runtime, without needing to tie the pipeline to a vendor. The move mirrors what VMware did in the 2010s with hypervisors: monetize control, not the engine.

Implications for Operations Outside the U.S.

For Brazilian, Indian, and European teams using AWS, the practical point is the native Web Search within AgentCore. The functionality had previously been accomplished through loose integrations, each with its own costs and governance, and now consolidates under the same guardrail and billing. For the Acceleration Centers of PwC and Deloitte in India, and for the shared service centers of Capgemini and TCS in São Paulo and Buenos Aires, the disappearance of the homemade orchestration layer removes justification for billed hours in "plumbing."

The more uncomfortable read is for EY, KPMG, Deloitte, and Accenture, which sell their own implementation of "agentic platforms." When the runtime is a commodity sold in GA with SLA, the value charged needs to be at another layer, vertical, or data, not in the foundation. Pressure on margins should appear in the upcoming earnings reports from these firms, first in the transformation and digital lines, still in this fiscal year.