September 2026 Summaries

AI agents that act on users’ behalf should use delegated, per-user OAuth 2.0 credentials with authorization code flow and PKCE whenever possible, ensuring that agents inherit only the permissions of the individual they represent and preserve clear audit attribution. Credentials should remain outside model prompts, tool inputs, and responses, with infrastructure attaching tokens only at execution time to prevent exposure through prompt injection, logging, or model providers. Security boundaries should be enforced through narrowly scoped grants rather than system prompts, while credentials should be encrypted, isolated by user, connector, and tenant, and never stored in process-wide environment variables. The guidance also recommends short-lived access tokens, server-side refresh handling with rotation and concurrency safeguards, and distinct handling for expiration versus revoked access, including pausing work and alerting users when permissions are removed. Teams should test failure scenarios such as expired tokens, revoked grants, missing scopes, and cross-user credential requests to ensure agents fail visibly and safely, while shared service accounts should be reserved for limited, low-sensitivity team use cases.

Merge has introduced the Universal Context Layer in Merge for Workforce, a system designed to give AI tools and agents secure, reusable access to an organization’s knowledge, data, and skills while operating under each employee’s existing permissions. The platform syncs and vectorizes information from connected file storage, ticketing, and HRIS systems, enabling agents to search a pre-indexed knowledge base, cite sources, identify relevant tools, and avoid repeated live API calls or manual context setup. Merge says this approach can improve AI output quality, reduce token consumption, and let organizations configure context once for use across multiple AI tools and agent platforms. The system also creates a feedback loop by analyzing usage patterns and presenting administrators with recommendations for relevant skills, cost-saving changes, and potential data fixes, which they can approve or decline.