OVERVIEW
Observe, control, secure, and govern AI agents across your organisation
- Foundry Control Plane is built for agentic systems. Trace agent runs end to end, including inputs, reasoning steps, tool calls, outputs, latency, and cost. Run preproduction evaluations and continuous evaluations on production traffic to measure task adherence, tool call accuracy, safety, and quality. Set thresholds, trigger alerts, and move from ad hoc testing to continuous oversight as your agent fleet scales.
- Enforce runtime controls at every stage of agent execution. Configure intervention points across user input, tool calls, tool responses, and outputs to detect and block PII, indirect prompt injection, task misalignment, prohibited actions, and content risks. Apply and update controls centrally so every agent reflects your organisation’s standards.
- Treat agents as first-class identities in your organisation. Automatically assign Entra Agent ID so each deployed agent has a verified identity and governed access. Surface jailbreak attempts and threat signals in Microsoft Defender for investigation and response. Extend Microsoft Purview data security and compliance policies to AI interactions, so agents operate with the same protections as users and devices.
- Run your AI operations from a single control surface. See active alerts across security, policy, cost, and performance in one prioritized view. Track agent runs, success rates, and spend in real time. Monitor compliance status, behaviors prevented, and usage trends week over week. View and manage agents, models, and tools across projects so you know what needs attention and where to act next.
FEATURES
Built to run agents in production
Operate AI agents with continuous evaluation, end-to-end tracing, runtime controls, and integrated security. Use a single control plane to monitor, enforce, and troubleshoot at scale.
Continuous evaluation on production traffic
Run continuous evaluations on production traffic to measure quality, safety, and task completion. Set thresholds and surface regressions automatically.
Agent-specific evaluators for real-world risk
Measure task adherence, tool call accuracy, intent resolution, groundedness, and sensitive data exposure. Detect the failure modes traditional language model evaluations miss in agentic systems.
End-to-end tracing with OpenTelemetry
Trace every agent run from inference through tool calls using OpenTelemetry standards. Integrate with Azure Monitor and Application Insights to diagnose drift, failures, and performance bottlenecks.
Tool-level runtime guardrails
Apply guardrails, not only for prompts and outputs, but also for tool calls and tool responses. Define risks, intervention points, and response actions where agents actually take action.
Protection against direct and indirect prompt attacks
Mitigate direct and indirect prompt injection attacks before agents act. Use advanced detection and spotlighting to uncover cross prompt manipulation and block malicious instructions in real time.
Task adherence to prevent agent drift
Keep agents aligned to intended tasks with runtime task adherence controls. Detect off-task behavior and enforce intent boundaries across inputs, tool calls, and outputs.
AI Red Teaming Agent
Simulate adversarial attacks and automatically scan agents for vulnerabilities before deployment. Identify systemic weaknesses with automated clustering and readiness reports.
Microsoft Entra Agent ID
Assign every agent a durable identity at build time. Apply policy-based access controls, conditional permissions, and lifecycle governance with Microsoft Entra for centralized oversight.
Proactive alerts across your agent fleet
Surface evaluation, policy, and security alerts with context and deep links into traces and remediation workflows. Move from reactive monitoring to proactive agent operations.