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.
By Hypler editorial team · Source: Microsoft.
Original source date: .
Hypler publication: . Hypler updated: .
Estimated reading time: 4 minutes.
Topics: Microsoft, Azure AI Foundry, agent service, platform

GA is a product stage, not an application verdict
Microsoft announced general availability of Azure AI Foundry Agent Service on May 29, 2025 in a post about software-company tooling around Microsoft Build. It described Foundry as a unified platform of models, tools, and services, with a unified API and SDK for orchestrating multi-agent workflows from existing development environments, including GitHub and Visual Studio. General availability gives a clearer vendor release record for the capability announced in May 2025. It does not replace decisions about models, instructions, data, tool access, evaluation, incident response, or authority for external effects in systems built on the service.
A unified platform still has separate owners
A managed agent service can supply model access, tools, identity, observability, and hosting. The application team owns the task contract: inputs, eligible source records, permitted tools, configuration authority, and output evaluation. Write those choices as versioned application state rather than leaving them implicit in a portal. In multi-agent work, give each component a role, input schema, output schema, allowed tools, and handoff condition. A coordinator may route a task but should not hide which agent produced a claim or requested an action. Clear interfaces permit replacement and investigation without treating the workflow as one opaque unit.
An incident draft is different from incident action
Consider a hypothetical internal incident-triage workflow. One agent summarizes a monitored alert, another retrieves approved runbook passages, and a third creates a draft incident note. Each receives structured input and returns typed output with citations. The system may create a draft in an incident workspace, while the on-call engineer owns diagnosis, escalation, and corrective action. Missing monitoring data, an outdated runbook, or disagreement between agents becomes a recorded conflict routed to the engineer. The coordinator does not invent a resolution or call a remediation tool. Evidence gathering remains useful without delegating operational authority.
Version the workflow as a graph of contracts
Store agent definitions, model choices, tool allowlists, schemas, routing rules, fixtures, and release notes together. Every execution identifies workflow and component versions. This lets teams compare a changed result with the configuration that produced it, rather than guessing whether model, prompt, tool, routing, or source data caused it. Handoffs should carry only information required by the next component. A research agent can pass claims and citations to a writer, while the writer lacks credentials and broad store access. A state-changing tool should receive a validated execution request, never unbounded conversational text.
Orchestration adds distributed-systems failures
One component can present a guess as established context for another. Tasks can loop without a terminal condition, branches can race on shared state, and tool calls can succeed after coordination timed out. Test duplicate delivery, expired credentials, model timeout, conflicting results, malformed handoffs, tool denial, and partial external action. Define completed, failed, cancelled, approval-required, and unknown terminal states. A trace should show decision order and inputs behind each transition. That record allows a team to investigate a failure without treating an agent's final message as proof of what every connected system did.
A release should include operating evidence
Evaluate the complete workflow on normal execution, incomplete evidence, invalid handoffs, tool failures, and reviewer rejection. Check factual support, schema conformance, tool policy, latency, cost, and audit-trail clarity. Use fixed fixtures and record versions of models, agents, tools, and policies. First production use should be bounded, owned, monitored, reversible, and backed by a manual fallback. Set thresholds for intervention, completion quality, unexpected tool calls, and unresolved outcomes. Revisit the workflow whenever a model, connector, agent definition, or routing policy changes. Hypler's judgment is that GA does not remove these responsibilities.
Hypler analysis
Hypler’s view
General availability establishes a vendor lifecycle milestone, not readiness for a particular agent system. Application teams still own the contracts, evaluations, cost limits, incident paths, and authority boundaries that turn platform capability into reliable work.
Hypler technical analysis
Engineering notes
A unified agent API and SDK can reduce integration boundaries, while model choice and routing make runtime behavior a controlled variable. General availability describes a product stage; operational readiness still depends on versioned evaluation, model and tool policy, cost controls, incident handling, and change evidence for each workflow.
What to inspect
- Route decision
- Record selected model or route, workload class, fallback behavior, invocation settings, tool access, and the owner who approved the selection.
- Operational limits
- Inspect request budgets, concurrency and retry rules, tool-call limits, stop conditions, and alerts for unexpected usage or repeated failures.
- Incident path
- Exercise a wrong model selection, unavailable dependency, unsafe tool request, and partial external outcome; inspect rollback, evidence capture, and escalation ownership.
Scope and limits. General availability does not establish that a specific agent architecture is fit for an organization's data, workflow, cost, safety, or incident requirements. Routing changes can create behavior changes even where application interfaces remain stable.
Terms and definitions
- Evaluation
- A repeatable method for measuring a system against defined quality, safety, or task criteria.
- 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.
- 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.
- Microsoft previews Microsoft Agent Framework
Microsoft announced the public preview of Microsoft Agent Framework, an open-source SDK and runtime for multi-agent orchestration.
- AWS previews Amazon Bedrock AgentCore
AWS previewed Amazon Bedrock AgentCore, modular services for runtime sessions, memory, gateways, browser use, code interpretation, identity, and observability.

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.