AI Platforms
[Update] Foundry Agent Optimizer Scores Hosted AI Agents
Published June 04, 2026 by Dillip Chowdary
Microsoft Foundry Agent Optimizer gives hosted agents a managed improvement loop: evaluate the current behavior, generate better configurations, score them, and promote the winner.
What Changed
- Preview timing: Microsoft introduced the feature on June 3, 2026 in private preview, with public preview planned in 30 days.
- Closed loop: The optimizer runs baseline evaluation, candidate generation, candidate evaluation, ranked recommendation, and deployment as one workflow.
- Optimization targets: Teams can tune instructions, generate reusable skills, or compare model deployments for quality and cost.
Architecture Impact
For engineering teams, the important shift is that agent quality management is moving from manual prompt editing into an observable release process. The source task set, pass/fail criteria, token cost, and candidate ranking become production artifacts rather than notes in a chat thread.
That changes how teams should store agent configuration. Keep prompts, skills, evaluation tasks, and deployment metadata versioned next to code so a promoted candidate can be reviewed, rolled back, and compared against the previous production agent.
Rollout Checklist
Start with one hosted agent and build a small task suite with explicit pass/fail criteria. Require approval before applying optimized instructions to production, and track score movement alongside latency, token cost, and safety regressions.
Do not optimize against happy-path examples only. Include refusal, escalation, missing-data, policy-boundary, and tool-failure scenarios before trusting the ranked candidate.