Agents, Tools & Digital Infrastructure

Evidence statusPrimary source · Technical preview

Claude Tag brought a delegated agent into selected Slack channels, where teams can assign tasks and connect approved tools, data and codebases. Administrative controls include spend limits and activity logs, making team-level agent governance part of the product design.

Evidence boundaryThe launch was a beta for eligible Team and Enterprise customers, and the cited internal adoption rate is an Anthropic-reported metric.

Evidence statusPrimary source · Research

OpenAI described a new system for synthesizing long-term ChatGPT memory with attention to freshness, continuity and relevance. The work treats memory maintenance as an active process, an important component for agents that must preserve context across months or years.

Evidence boundaryThe rollout was staged by plan and geography; persistent memory raises privacy, correction and user-control requirements that must be assessed independently.

Evidence statusPrimary source · Research

Anthropic outlined containment patterns used across Claude products to cap the potential blast radius of increasingly capable agents. The engineering approach layers isolation, permission boundaries, approval gates and monitoring rather than relying on model behavior alone.

Evidence boundaryThe article describes Anthropic's own architecture and lessons; it does not provide full implementation details or an external assurance assessment.

Evidence statusPrimary source · Product release

Google added managed agents to the Gemini API, providing isolated Linux execution, resumable state and file-defined instructions and skills. The Antigravity harness lets developers launch a tool-using coding agent or define custom agents through the Interactions API.

Evidence boundaryManaged execution reduces infrastructure work but does not remove the need for application-specific authorization, testing, observability and cost controls.

Evidence statusPrimary source · Product release

OpenAI extended the Agents SDK with a model-native harness and managed sandbox execution so agents can inspect files, run commands, edit code and continue long-horizon work in controlled environments. The architecture separates orchestration from compute for security and durability.

Evidence boundarySafe operation still requires scoped credentials, tool policies, monitoring and application-level approvals; sandboxing does not eliminate all agent risk.

Evidence statusPrimary source · Research

Anthropic translated its trustworthy-agent principles into concrete controls across four layers: model, harness, tools and operating environment. The analysis highlights human control, scoped permissions, prompt-injection resistance, transparency and privacy as system-level requirements.

Evidence boundaryThis is Anthropic's governance framework and product interpretation, not a consensus standard or independent certification.

Evidence statusPrimary source · Standard

NIST's initiative focuses on interoperable protocols, agent security, identity and authorisation, and community-led open-source infrastructure.

Evidence statusPrimary source · Product release

OpenAI Frontier was introduced as an enterprise layer for building, deploying and governing AI agents across company systems. Its emphasis is operational infrastructure—identity, context, evaluation, permissions and lifecycle management—rather than a single assistant interface.

Evidence boundaryThe page includes selected enterprise outcome claims supplied by OpenAI and customers; organizations need their own controlled pilots and measurement.