The AI Wire

Lead dispatchtechcommunity.microsoft.com ▸ azure· filed 2 Oct 2026

Federated Agent Factory: Microsoft's Reference Architecture for AI Agents Across Entra Tenants

Microsoft's reference architecture for multi-tenant AI agent governance uses a federated agent factory, keeping agent identities and data within business-unit tenants while a governing tenant sets standards and operational views. It specifies runtime enforcement via Azure API Management for pro-code agents and Agent 365/Purview for Copilot Studio agents, with policy-as-code for per-action authorization. Several components remain in preview, including the dedicated AI Gateway tier and multitenant agent management.

Read the dispatch →

No. 087

Latest on the wire37 dispatches · page 01 / 04

techcommunity.microsoft.com ▸ models· filed 2 Oct 2026

Microsoft adds streaming transcription and two text-to-speech models to Azure AI Foundry

Microsoft has introduced three speech models to Azure AI Foundry: one that streams partial transcripts before a speaker finishes, and two text-to-speech models that trade off latency and cost. The streaming model is priced at $0.54 per hour until the end of the year, while the voice models run $22 and $15 per million characters. Benchmark comparisons come from vendor-reported figures, and the pricing of transcription for the long term remains unspecified.

github.blog ▸ product· filed 27 Sept 2026

GitHub's HydraFusion preview orchestrates coding models at runtime

GitHub has released Project HydraFusion as a research preview in Copilot, a runtime orchestrator that selects among single-model, cascade, and critique-revise workflows for each coding request. Offline evaluations against Claude Opus 5 on three benchmarks show mixed quality—one gain, two slight shortfalls—with estimated costs down 36–67%, based on GitHub's own assumptions rather than production measurements. The preview is limited to well-scoped first-turn tasks in autopilot mode, with multi-turn sessions listed as the next focus.

azure ▸ product· filed 10 Sept 2026

Claude Fable 5.1 arrives in Microsoft Foundry alongside restricted Mythos 5.1

Anthropic has released Claude Fable 5.1 in the Microsoft Foundry model catalog and restricted Claude Mythos 5.1, describing both as the same underlying model with different safeguard configurations. The company says the main advance is sustained performance on multi-step, tool-use workflows, with fewer overconfident errors and better reporting of incomplete work. Anthropic recommends testing on realistic workloads and simplifying prompts built for earlier models, though these claims have not been independently verified.

azure ▸ development· filed 6 Sept 2026

Microsoft Foundry details six-phase process for migrating LLM models

Microsoft Foundry has detailed a six-phase process for moving LLM-powered applications between model generations, with model retirement dates now built into the platform lifecycle. The guidance emphasizes that swapping model names can silently change response formatting, schemas, and tool behavior, and that validation and canary rollouts are necessary because offline evaluation can miss production issues. A worked example ends in a deliberate split deployment rather than a clean cutover.

simonwillison.net ▸ product· filed 31 Aug 2026

Simon Willison's hands-on test of ChatGPT Work: open internet access, browser automation, and unanswered safety questions

Simon Willison's hands-on test shows ChatGPT Work is an agent platform with unrestricted internet access, a headless Chrome browser, and parallel sub-agents, but OpenAI has not explained how it defends against prompt injection. The documentation is so sparse that he had to ask the agent itself to list its 223 tools and 44 skills. This leaves core safety mechanisms unaddressed in OpenAI's marketing.

youtube.com ▸ models· filed 30 Aug 2026

GLM 5.3 Flash: Z.ai’s cheaper, multimodal sibling matches larger models but slows on max reasoning

Z.ai's GLM 5.3 Flash is a 320-billion-parameter multimodal model that costs about one-ninth the output price of the flagship GLM 5.3. Independent tests show strong function-calling and agentic behavior, but max‑reasoning mode slows throughput to 50–60 tokens per second, and agentic fallbacks to the larger GLM are expensive, pushing users toward other frontier models for critical tasks.

youtube.com ▸ industry· filed 30 Aug 2026

Andrew Ng: Big AI labs are fear-mongering to shape regulation

Andrew Ng contends that leading AI companies are exaggerating risks to drive regulation that favors incumbents, while dismissing fears of a job apocalypse and warning that AI models hinder learning. He advises individuals to adopt AI skills and focus on product decisions rather than fear.

azure ▸ product· filed 30 Aug 2026

Grok 4.6 available in Microsoft Foundry public preview

SpaceXAI's Grok 4.6 is now available in Microsoft Foundry's public preview, aimed at long-horizon agentic work like repository issue resolution and business deliverables. Provider-reported benchmarks show strong performance on terminal and 3D modeling tasks, but production reliability remains unproven outside controlled evaluations.

azure ▸ product· filed 30 Aug 2026

Cohere Parse Document Extraction Model Now Available in Microsoft Foundry

Cohere Parse, a vision language model for extracting structured data from documents, is now available in Microsoft Foundry on Azure, converting tables, forms, and embedded images into Markdown while keeping inference within the customer's tenant. The model returns bounding boxes for downstream citation and reports throughput of 4.5 pages per second per GPU at $1.50 per thousand pages. Its domain-focused training on finance, insurance, scientific research, and nine languages means teams with other document types should validate output on their own corpus.

developer.nvidia.com ▸ research· filed 22 Aug 2026

AVO agent architecture achieves perfect ARC-AGI-3 score and produces optimized GPU kernels

The AVO agent architecture, designed for long-running autonomous work, achieved a perfect score on the ARC-AGI-3 public set and produced GPU kernels that outperformed existing libraries. The system uses persistent memory and a supervisor to sustain progress. Results show fewer environment actions than a comparable harness, but limitations include public-set-only evaluation and unqualified cross-model comparisons.