How long does a Microsoft Foundry agent build take?

The question usually arrives right after someone watches a Foundry demo and needs a date for the roadmap. Scope is the real variable, and scope is a decision the buyer makes. The platform is rarely the constraint.

For a CIO or VP of IT, the number has to survive a steering committee, so it helps to know what it's made of. An agent that answers questions from a governed document set is a different piece of work from an agent that opens tickets in a service desk, and both get called "an agent" in the same meeting. Microsoft's Cloud Adoption Framework splits agent adoption into four phases, and building is only one of them (Microsoft Learn, "AI agent adoption," accessed September 1, 2026). Most schedule risk sits upstream of the build, in whether the data estate can be queried under existing permissions without a remediation project first. The tiers below are how ArchitectNow turns that risk into a date.

The three scope tiers ArchitectNow plans against

The tiers describe ArchitectNow's planning ranges for scoping and staffing a build. They aren't customer results, and a real engagement gets dated only after a scoping pass on the specific environment.

What decides which tier a Foundry agent build lands in?

Three inputs set the tier. The first is how many systems the agent has to reach. The second is whether the agent writes back to any of them or only reads. The third is the condition of the data behind those systems. Model selection and agent type barely move the date by comparison.

Foundry itself gives you three build paths, and the choice between them costs less time than most teams expect. A prompt agent is defined entirely through configuration, down to the instructions and the tool set, and Foundry runs it with no code or infrastructure to manage. A hosted agent is your own code and framework, shipped as a container image or a .zip file, which Foundry runs behind a managed endpoint with automatic scaling plus a dedicated Microsoft Entra identity. The Responses API is a third option for teams that already run agent code elsewhere and want Foundry models and platform tools without a persisted agent resource (Microsoft Learn, "Agents in Microsoft Foundry," accessed September 1, 2026).

Picking among those three is an afternoon of architecture. What sets the calendar is the surface area behind them. A read-only agent over a single curated index has one integration and one failure mode. An agent that creates a ticket adds an authentication decision, an input validation layer, a rollback story, then a named human who owns the consequences of a wrong write.

Microsoft's guidance is explicit that tools change the risk profile: the moment an agent moves from retrieving information to engaging systems, it can modify data and trigger workflows, which is why the Cloud Adoption Framework calls for narrowly scoped endpoints and human-in-the-loop confirmation on high-impact actions (Microsoft Learn, "Process to build agents across your organization," accessed September 1, 2026). Every one of those controls is real work with a real duration.

What does a two-week Foundry agent look like?

A two-week build is a Foundry prompt agent grounded in one knowledge source, answering questions and citing where the answer came from. It writes nothing. The content set already exists and already carries the right permissions. One person can say what belongs in it. Scope discipline is what keeps this at two weeks.

The mechanics are well documented. The Azure AI Search tool retrieves indexed documents so the Foundry model powering the agent can generate answers with inline citations (Microsoft Learn, "Connect an Azure AI Search index to Foundry agents," accessed September 1, 2026). A SharePoint connection covers the same job when the authoritative content already lives in Microsoft 365.

Inside two weeks the work divides roughly in half. The first half is content and permissions: confirming the source of truth, indexing it, then testing that retrieval respects the access model people already have. The second half is behavior: writing the agent instructions, running an evaluation set against known-good answers, and tightening the instructions until groundedness holds.

That evaluation step is what separates a two-week agent from a demo. Foundry ships built-in evaluators covering general quality metrics such as coherence and fluency, retrieval metrics such as groundedness and relevance, plus agent-specific metrics like tool call accuracy and task completion (Microsoft Learn, "Observability in generative AI," accessed September 1, 2026). Running them is cheap. Skipping them is how an agent ships confident and wrong.

One caution: connecting an index doesn't guarantee the model uses it. Instructions have to mandate the tool and forbid answering from general knowledge, which is why the Cloud Adoption Framework treats tool mandates and citation rules as required parts of an instruction architecture.

What does a two-month Foundry agent look like?

A two-month build is a Foundry hosted agent running custom orchestration across several connected systems, with at least one write path into a system of record. It carries its own repository, deployment pipeline, evaluation suite, plus a production support model. The extra weeks buy governance and reversibility.

Hosted agents let a team bring code written with the Microsoft Agent Framework, LangGraph, the OpenAI Agents SDK, or their own stack. Foundry runs it behind a managed endpoint with session-level state persistence plus a dedicated Entra identity per agent (Microsoft Learn, "Agents in Microsoft Foundry," accessed September 1, 2026). That identity matters more than it sounds. It's what makes least-privilege access and per-agent auditing possible once several agents are running.

The write path is where the weeks go. Foundry supports key-based access, Microsoft Entra Agent Identity, project managed identity, or OAuth identity passthrough for tool calls, and Microsoft recommends identity passthrough when user-level permissions have to persist through the agent (Microsoft Learn, "Data architecture for AI agents," accessed September 1, 2026). Choosing among those, wiring them, then proving them in a sandbox that mirrors production is a multi-week exercise on its own.

Then there's everything that has to be true before go-live. Traces flowing into Application Insights. Continuous evaluation sampling live traffic. Guardrails applied at input, at the tool call, at the tool response, then at the final output. An AI red teaming pass against prompt injection. Microsoft frames these as layered controls precisely so that no single failure point takes the system down, and layered controls take longer to build than a single checkpoint.

What is the first week of a Foundry agent build spent on?

The first week is spent deciding what the agent is not allowed to do. Microsoft calls this the agent charter: a governance artifact documenting the agent's responsibilities, its role boundaries, plus the actions it must never take. Teams that skip it pay for it later in scope creep and compliance rework.

That week produces four things a build can't proceed without. A charter that maps the agent's job to a business objective and names what it must avoid. An instruction architecture covering identity and tone, scope and boundaries, tool mandates, plus citation rules. A retrieval decision per data domain, classified as search, API, or both. Finally, an owner for every action the agent can take.

The retrieval documentation is the piece most teams underestimate. Microsoft's guidance is to list each data domain the agent supports and classify how the agent reaches it, with examples as concrete as "Product FAQs: search using a knowledge index" and "Order tracking: API call to the order management system" (Microsoft Learn, "Data architecture for AI agents," accessed September 1, 2026). Writing that list out is usually the moment a team discovers that two of its five domains have no clean access path, which is exactly the discovery you want in week one.

ArchitectNow runs this pass before any build date is committed. In one AI Day workshop, the intake produced 131 AI use cases, consolidated into 18 thematic opportunity areas and 5 ranked should-do initiatives (ArchitectNow AI Day workshop output). The ranking mattered more than the volume. Tier 1 candidates had been sitting on the same list as Tier 3 ambitions, undifferentiated.

When is Copilot Studio the faster path than Foundry?

Copilot Studio is faster when the agent serves a department, connects through existing Power Platform connectors, and doesn't need custom orchestration. Microsoft positions it for small-scale deployments such as department-level solutions, built by makers through a drag-and-drop interface. For that shape of problem it beats a Foundry build on time to first value.

The trade is developer control. Microsoft's own comparison names the limits plainly: less ability to fine-tune API calls, actions, or response formatting compared with a pro-code environment, and pro-dev tooling that's less integrated than a project-based workflow (Microsoft Learn, "Choose the right tool to build your declarative agent," accessed September 1, 2026). Governance, on the other hand, is a strength. Copilot Studio agents inherit environment-level data loss prevention policies, role-based access, plus auditing managed from the Power Platform admin center.

The two platforms also compose. A Copilot Studio agent can call a Microsoft Foundry agent, which lets a low-code front door sit over a pro-code reasoning engine. One prerequisite catches teams out: the Foundry agent must have the Activity protocol endpoint enabled, because a newly created Foundry agent exposes only the Responses and A2A endpoints by default, and the connection fails at runtime with a 400 error without it. That endpoint can currently be enabled only programmatically through the REST API or the Python SDK, and the Foundry portal keeps showing only Responses and A2A afterward (Microsoft Learn, "Connect to a Microsoft Foundry agent (preview)," accessed September 1, 2026).

Budget half a day for that detail if the architecture crosses both platforms. It's what turns a two-week estimate into a three-week one.

What stretches every Foundry agent timeline?

Data readiness. It's the one step that lands on every tier, and it's the step that turns a two-week estimate into a two-month project when the answer comes back wrong. The platform work is predictable. The state of the data estate is what varies between one organization and the next.

Microsoft states the dependency directly: "Because agents synthesize information rather than create it, their accuracy depends entirely on the quality and accessibility of underlying sources. Fragmented or ungoverned data leads to misleading results and security risks" (Microsoft Learn, "Data architecture for AI agents," accessed September 1, 2026). That sentence is the whole timeline question in two lines.

In practice, readiness comes down to four checks.

1. Does an authoritative version of the content exist, or do four copies exist with no owner?
2. Do the permissions on that content reflect who should actually see the answers?
3. Can the agent reach it through a built-in retrieval capability such as Foundry IQ, Fabric IQ, a Fabric data agent, an Azure AI Search index, or a SharePoint connection?
4. Is there a governance layer, typically Microsoft Purview, already classifying what's sensitive?

Four yeses and the build tracks its tier. One no and remediation goes in front of the build, which is where the extra weeks come from. This is also why ArchitectNow's fixed-scope pilot offers are structured at four weeks with a design specification in week one: the shape of the work is known, so the variable is the environment it lands in.

Remediation is rarely wasted work. A governed content set built for one agent is the foundation for the next one, so the effort carries forward.

The scoping pass that sets the date

A useful date comes out of one working session. Walk the intended use case through the tier table, name every system the agent reaches, mark which ones it writes to, then run the four data readiness checks against each source. That produces a tier, a range, plus a short list of what has to be true before the clock starts.

Product naming note for anyone searching older material: Microsoft's AI platform evolved from Azure AI Studio to Azure AI Foundry to Microsoft Foundry, which is the current name. The AI services portfolio moved from Azure Cognitive Services to Azure AI Services to Foundry Tools. Two dates are worth having on the calendar as well, since both fall on August 26, 2026: the azure-ai-inference package retires, and the Assistants API sunsets in favor of the generally available Foundry Agent Service.

Every ArchitectNow offer opens with a 30-minute sponsor call, and that call is the right place to run the scoping pass. ArchitectNow designs and delivers AI and cloud solutions on the Microsoft stack, holds four Microsoft Solutions Partner designations, and carries the AI Apps on Microsoft Azure Specialization. Reach the team at info@architectnow.net or (314) 366-0875.

Sources and references

1. Microsoft Learn, "Agents in Microsoft Foundry." https://learn.microsoft.com/azure/foundry/agents/overview (accessed September 1, 2026)
2. Microsoft Learn, "Migrate from the Foundry (classic) portal." https://learn.microsoft.com/azure/foundry/how-to/navigate-from-classic (accessed September 1, 2026)
3. Microsoft Learn, "Data architecture for AI agents." https://learn.microsoft.com/azure/cloud-adoption-framework/ai-agents/data-architecture-plan (accessed September 1, 2026)
4. Microsoft Learn, "Process to build agents across your organization." https://learn.microsoft.com/azure/cloud-adoption-framework/ai-agents/build-secure-process (accessed September 1, 2026)
5. Microsoft Learn, "AI agent adoption." https://learn.microsoft.com/azure/cloud-adoption-framework/ai-agents/ (accessed September 1, 2026)
6. Microsoft Learn, "Observability in generative AI." https://learn.microsoft.com/azure/foundry/concepts/observability (accessed September 1, 2026)
7. Microsoft Learn, "Connect an Azure AI Search index to Foundry agents." https://learn.microsoft.com/azure/foundry/agents/how-to/tools/ai-search (accessed September 1, 2026)
8. Microsoft Learn, "Choose the right tool to build your declarative agent." https://learn.microsoft.com/microsoft-365/copilot/extensibility/declarative-agent-tool-comparison (accessed September 1, 2026)
9. Microsoft Learn, "Choose between Microsoft 365 Copilot and Copilot Studio to build your agent." https://learn.microsoft.com/microsoft-365/copilot/extensibility/copilot-studio-experience (accessed September 1, 2026)
10. Microsoft Learn, "Connect to a Microsoft Foundry agent (preview)." https://learn.microsoft.com/microsoft-copilot-studio/add-agent-foundry-agent (accessed September 1, 2026)
11. Microsoft Learn, "Agent identity concepts in Microsoft Foundry." https://learn.microsoft.com/azure/foundry/agents/concepts/agent-identity (accessed September 1, 2026)
12. ArchitectNow AI Day workshop output (anonymized), claims library entry stat-ai-day-outputs.