Today's AI trends
AWS’s September 8 reference implementation puts agent behavior directly into CI/CD: it deploys an agent and MCP server, invokes evaluation prompts, scores the results, and blocks pull requests on regressions. AWS’s implementation details are here. The practical implication is to treat representative agent scenarios as a release artifact, not a demo script.
The OpenAI Agents SDK documents Guardrails before and after guarded function Tools and local MCP Tool calls. Its MCP guidance also recommends trusted servers, least-privilege Credentials, keeping tokens out of URLs, and approval for sensitive operations. Read the Guardrails documentation and the MCP guidance. The operational shift is toward enforcing policy at the action boundary, where side effects occur, rather than relying only on a Prompt or final-output filter.
GitLab documents connecting Duo Self-Hosted to GPT, Claude, Llama, and Mistral models hosted in Microsoft Foundry, with administrators able to assign models by Duo feature. GitLab’s September 8 guide describes the setup. For regulated teams, model choice is therefore becoming a platform design question involving identity, network isolation, residency, compatibility, and workload routing—not merely a chat setting.