Microsoft previews Deep Research in Azure AI Foundry
Microsoft announced a public preview of Deep Research in Azure AI Foundry Agent Service. The company described an API and SDK offering based on OpenAI's research capability, with Bing Search grounding and source-backed output.
By Hypler editorial team · Source: Microsoft.
Original source date: .
Hypler publication: . Hypler updated: .
Estimated reading time: 4 minutes.
Topics: Microsoft, Azure AI Foundry, research, agents

The preview described a composable research tool
Microsoft announced Deep Research in Azure AI Foundry Agent Service on July 7, 2025. It described an API and SDK offering based on OpenAI's research capability, grounded with Bing Search and intended to return source-backed output. Microsoft presented it as a composable tool for applications, workflows, or agents, with Foundry connectors, Logic Apps, and Azure Functions integrations. It said the feature was available in limited public preview for Foundry Agent Service customers. The article described scoping a task, gathering web information, executing research, and returning a cited structured report with clarifications.
A research request needs a decision context
Define the question, allowed sources, date range, jurisdiction or market where relevant, audience, desired evidence type, and unanswered questions that must remain visible. These inputs keep web retrieval from becoming a generic summary and give reviewers a basis for judging whether citations address the assigned task. Citation traceability does not establish source quality; recency, authorship, primary evidence, and claim-to-source relationship still matter. Return claim list, citation for each claim, source dates, retrieval time, and uncertainty labels. The result becomes an artifact that can be inspected, revised, or declined before it affects a business decision.
A regulatory brief can preserve uncertainty
Imagine a hypothetical product team researching a proposed regulation. Its request names jurisdiction, date cutoff, preferred primary sources, and decisions it expects to make. The agent gathers public material and returns a draft with section citations, publication dates, and questions where the record is incomplete. It has no access to customer data, internal contracts, or publication channels. A policy reviewer compares the draft with named sources. A moved source, wrong jurisdiction, or conflict among primary records remains highlighted rather than smoothed into a confident conclusion. Research becomes reviewable input, not delegated policy judgment.
Research, synthesis, and action are different stages
Collection should query approved sources and capture provenance. Synthesis should organize claims, citations, and unresolved questions. Downstream action should receive a reviewed artifact and follow its own policy for changes to systems of record. Structured state can hold request identifier, source-policy version, query-log reference, result version, reviewer, and terminal status. A reporting connector should receive an approved brief identifier and approved destination, rather than the entire web-search session. This reduces data movement and leaves each transition inspectable if a report requires correction or a source later proves unsuitable.
Confident synthesis can hide coverage gaps
Research automation can miss a primary record, overweight a ranked secondary page, use stale material, or combine claims from different dates and jurisdictions. Design reports to show source date, source type, direct support for material claims, and open questions. Test ambiguity, conflicting sources, unavailable pages, instructions embedded in a retrieved page, and a request for a conclusion beyond evidence. Check whether citations support the precise claim and whether research is distinguished from recommendation. These cases matter even when common summaries are accurate because they reveal the boundary that matters for high-stakes use.
Measure evidence quality and review cost
Build tasks with independently reviewed source sets and known uncertainty. Score factual support, citation accuracy, coverage, stale and conflicting-information handling, and usefulness of the unresolved-questions list. Measure reviewer verification time as well as generation time; a fast report that is expensive to check is a poor operating trade. Start with one recurring research task, define source policy, and keep publication and external communication outside automation. Retain configuration, input policy, retrieval date, review outcomes, and corrections. Re-run the benchmark after source, tool, prompt, or routing changes. Preserve the benchmark's decision context alongside its scores.
Hypler analysis
Hypler’s view
Citations make a research output inspectable, not automatically sufficient. A reusable research tool needs a defined question, source policy, evidence standard, and human decision point before its polished synthesis influences a consequential choice.
Hypler technical analysis
Engineering notes
A research capability exposed through an API becomes a reusable tool whose invocation rules, search grounding, source selection, and output review need a defined contract. Source-backed output helps inspection, but citations must be assessed against the question, source authority, extraction context, and permitted use.
What to inspect
- Invocation policy
- Inspect which workflows may invoke research, the allowed question class, budget or stop rule, source constraints, and conditions requiring human validation.
- Grounding record
- For a representative response, inspect query formulation, source identifiers, cited claim linkage, inaccessible or conflicting sources, and unsupported assertions.
- Review gate
- Test a high-consequence conclusion, an ambiguous question, and a stale source; inspect whether output is labeled for review rather than silently promoted to action.
Scope and limits. The preview's API, SDK, search grounding, and connectors do not guarantee source quality, complete coverage, or correct conclusions for any particular research task.
Terms and definitions
- Tool Calling
- A model interface pattern where the model requests a structured operation from an external tool.
- Least Privilege
- A security principle that grants only the minimum access needed for a specific task.
- Execution Evidence
- Records that show whether an approved action was attempted, what inputs and targets it used, and what result occurred.
- Approval-First Automation
- A workflow that requires review and approval before a consequential state-changing action runs.
- Microsoft makes Azure AI Foundry Agent Service generally available
Microsoft announced the general availability of Azure AI Foundry Agent Service, together with a unified API and SDK for building and orchestrating multi-agent workflows.
- AWS previews the Amazon Nova Act SDK
AWS previewed the Amazon Nova Act SDK for browser automation agents, including IAM, S3 data controls, a browser tool, Python scripting, and testing support.
- OpenAI introduces ChatGPT agent
OpenAI announced ChatGPT agent, a product feature that combined web interaction, research, code execution, and document work in a virtual computer environment.

Original source
Historical source briefing with Hypler engineering analysis. Source details reflect the original publication date; implementation examples are hypothetical. Technical analysis expanded October 24, 2026; this does not change the original source event or establish a new source verification. Editorial prose revised October 26, 2026; original event, publication and source-check dates are retained.