Azure Foundry Adds Voice Agents, Continuous Optimization, and 97% Token Reduction in Tool Search

Microsoft’s Azure Foundry platform has been accumulating agent-oriented capabilities across 2026. The latest update batch addresses three distinct problems teams encounter when moving agents from prototype to production: voice channel deployment, iterative performance improvement, and the token cost of giving agents access to large tool catalogs.

Voice Agents in Public Preview

Foundry Agent Service now supports voice as a first-class agent type. The implementation covers multiple voice model backends (GPT Realtime, Azure Realtime, MAI), handles 80+ languages and 140+ locales, and routes to web, Microsoft Teams, Teams Phone, and Twilio telephony endpoints from a single deployment configuration.

The significance of Teams Phone integration is specific to enterprises with existing Teams telephony infrastructure. Building a voice agent that handles incoming calls to a Teams Phone number — customer support queues, appointment scheduling, intake workflows — becomes a Foundry deployment rather than a custom telephony integration project.

Four Continuous Optimization Tools

The optimization layer addresses a common production agent problem: agents are tested before deployment but not systematically improved from production behavior afterward. All four tools go GA this month:

Insights in Foundry (public preview): Analyzes production traces to surface recurring failure patterns and recommended adjustments. Turns operational monitoring into an active improvement signal rather than post-mortem data.

Rubric Evaluator: Defines measurable success criteria for agent workflows as structured rubrics, enabling consistent evaluation against explicit standards rather than subjective assessment.

Traces to Dataset Generation: Converts production traffic into evaluation datasets automatically. Real agent interactions — including failures — become test cases without requiring manual curation.

Agent Optimizer: Tests configuration changes (instructions, tools, skills, model selection) systematically against the real production trace dataset. Eliminates the manual cycle of changing something, deploying, waiting for results, and interpreting outcomes.

The practical value of connecting these four tools is that the production feedback loop becomes automated: Insights identifies what’s failing, Traces captures the real-world cases, Rubric Evaluator defines what success looks like, and Optimizer tests changes against both. This closes the gap between “agent works in staging” and “agent improves over time in production.”

The 97% Tool Token Reduction

The most concrete number in this release: Foundry’s tool search mechanism reduces input-token consumption by more than 60% for a 50-tool toolbox and by over 97% for a 1,000-tool toolbox, compared to prompt-cached baseline.

The mechanism works by selectively surfacing relevant tools based on the current context rather than including all tool definitions in every prompt. For agents with small tool catalogs, the savings are meaningful. For agents with large catalogs — the kind that enterprise automation typically builds over time as more integrations get added — the reduction is dramatic.

A 1,000-tool catalog going from prompt-cached to 97% reduction in tool tokens changes the economics of large-tool-catalog agents significantly. The cost and latency of running complex enterprise automation agents drops in proportion to how many tool definitions were previously being loaded into context on every turn.

Model Availability

GPT-6 Sol and Luna (from OpenAI) and Claude Opus 5.5 (from Anthropic) are now available through Foundry alongside existing models. The optimization tooling is designed to test across model options — Agent Optimizer explicitly evaluates model selection as one of the tunable variables, so the expanded model catalog feeds directly into the optimization loop.

Source: Microsoft Azure Blog

Content created with AI assistance and reviewed for accuracy.