AI agents are moving quickly from proof-of-concept demos into real business workflows. For Australian organisations, the question is no longer whether agents are interesting. The question is how to build them in a way that is secure, supportable, cost-aware, and aligned with governance expectations.
Azure AI Foundry gives teams more than one way to build agents. Two important options are Prompt Agents and Hosted Agents. They both sit under the broader Azure AI Foundry Agent Service capability, but they serve different implementation needs.
Choosing the wrong option can create unnecessary complexity, slow delivery, or introduce operational risk. Choosing the right option can help teams move faster while keeping control over identity, data, observability, and compliance.
The short version
For most teams, the decision can be simplified like this:
- Use Prompt Agents when the agent can be defined mostly through instructions, model choice, and approved tools.
- Use Hosted Agents when the agent needs custom code, custom orchestration, specific runtime behaviour, or deeper integration patterns.
Prompt Agents are usually the faster starting point. Hosted Agents are usually the more flexible engineering option.
What are Prompt Agents?
A Prompt Agent is a declarative agent. It is configured using a model, instructions, and tools rather than being built as a custom application from the ground up.
In practical terms, a Prompt Agent is useful when the main behaviour can be described through prompts and configuration. A team can define what the agent should do, what tone or constraints it should follow, which tools it can use, and which data sources it can reference.
This makes Prompt Agents suitable for scenarios such as:
- Internal knowledge assistants
- Policy or procedure Q&A
- Document-based support assistants
- Simple workflow helpers
- Research or summarisation agents
- Service desk triage assistants
Prompt Agents are attractive because they reduce the amount of custom code required. This can help business and technology teams validate use cases quickly before committing to a larger engineering investment.
What are Hosted Agents?
Hosted Agents are code-based agents that run in a managed Azure AI Foundry hosting environment. Instead of defining the agent only through prompt configuration, developers package agent logic as an application, commonly using a container-based deployment model.
This gives the engineering team more control over the runtime, orchestration logic, state handling, external integrations, and framework selection.
Hosted Agents are better suited to scenarios such as:
- Custom multi-step orchestration
- Agents that need complex business rules
- Integration with internal APIs or line-of-business systems
- Custom protocols or request handling
- Advanced state management
- Multi-agent patterns
- Use of frameworks such as Semantic Kernel, LangGraph, or Microsoft Agent Framework
- Workloads that need more control over execution and error handling
Hosted Agents provide flexibility, but that flexibility comes with more design, development, testing, and operational responsibility.
When Prompt Agents make sense
Prompt Agents are a good fit when the business problem is clear and the workflow is not overly complex.
For example, an organisation may want an assistant that helps staff understand internal HR policies, IT procedures, product documentation, or compliance obligations. If the agent mainly needs to retrieve information, summarise it, and respond according to clear instructions, a Prompt Agent may be enough.
Prompt Agents also make sense when speed matters. Many AI initiatives fail because teams spend too much time building infrastructure before validating the use case. A Prompt Agent allows teams to test value quickly, collect feedback, and decide whether the workflow deserves further investment.
From a governance perspective, Prompt Agents can also be easier to standardise. Organisations can define approved models, approved tools, grounding patterns, and review processes before allowing wider adoption.
When Hosted Agents make sense
Hosted Agents become more relevant when the agent is no longer just responding with grounded information, but actively coordinating business logic.
Examples include:
- An agent that checks multiple systems before recommending an action
- A support agent that creates, updates, and tracks tickets
- A finance assistant that follows strict approval paths
- A compliance agent that applies deterministic policy checks
- A field operations agent that coordinates tasks across several APIs
In these cases, prompt configuration alone may not be enough. The organisation may need custom code to manage authentication, retries, validation, logging, exception handling, and business-specific orchestration.
Hosted Agents also give developers more freedom to use preferred frameworks and patterns. That can be important where the organisation already has established software engineering practices, CI/CD pipelines, testing standards, and application support models.
Security and compliance considerations
For Australian organisations, agent architecture needs to be assessed through a security and compliance lens from the beginning.
Relevant questions include:
- What data will the agent access?
- Is any personal information involved under the Privacy Act?
- Does the agent need to operate within specific data residency or network boundaries?
- How will access be controlled using Microsoft Entra ID and role-based access control?
- What logging and monitoring will be retained?
- How will the organisation detect misuse, prompt injection, or unsafe outputs?
- How does the solution align with ACSC guidance and Essential Eight maturity goals?
Prompt Agents can reduce custom application risk, but they still need strong controls around data access, grounding, identity, and monitoring.
Hosted Agents introduce more engineering flexibility, but also more responsibility. The code, dependencies, container images, secrets, APIs, and deployment pipeline all need to be secured. This makes practices such as vulnerability scanning, least privilege access, controlled releases, and audit logging especially important.
Cost and operational impact
Cost is another practical consideration.
Prompt Agents may reduce initial build cost because less custom development is needed. They can be a good way to test demand and usage patterns before scaling.
Hosted Agents may cost more to build and operate, but they can reduce long-term complexity when the use case requires proper application logic. Trying to force a complex workflow into a prompt-only pattern can become expensive in other ways: rework, support issues, inconsistent outcomes, and governance gaps.
The better question is not simply which option is cheaper. The better question is which option gives the right level of control for the business risk involved.
A practical decision framework
A useful way to decide is to ask five questions.
1. Is the workflow mostly conversational or procedural?
If the agent mainly answers questions, summarises content, or helps users navigate information, start with a Prompt Agent.
If the agent must execute a defined process with branching logic, validation, and system updates, consider a Hosted Agent.
2. How much custom integration is required?
If built-in or standard tool integrations are enough, a Prompt Agent may be suitable.
If the agent needs custom APIs, complex authentication, or business-specific orchestration, Hosted Agents are likely a better fit.
3. What level of control is required?
Prompt Agents trade some control for speed and simplicity.
Hosted Agents provide more control over code, runtime behaviour, state, and integration patterns.
4. What is the risk profile?
For low-risk internal productivity use cases, Prompt Agents can be a sensible first step.
For customer-facing, regulated, financial, operational, or safety-sensitive workflows, deeper engineering controls may be required.
5. Will this become a production application?
If the agent is a lightweight assistant, a Prompt Agent may remain the right model.
If the agent is expected to become part of a core business process, it may need the structure and lifecycle management of a Hosted Agent.
Recommended adoption path
For many organisations, the best path is not to choose one model forever. It is to mature through stages.
A common approach is:
- Start with a Prompt Agent to validate the business problem.
- Add approved data sources and tools.
- Measure usage, accuracy, cost, and support needs.
- Review security, privacy, and compliance requirements.
- Move to a Hosted Agent if the use case requires custom orchestration or production-grade application logic.
This staged approach helps avoid over-engineering early while still leaving room to scale properly.
Final thoughts
Azure AI Foundry Prompt Agents and Hosted Agents are not competing ideas. They are different implementation models for different levels of complexity.
Prompt Agents help teams move quickly with configuration-driven agent experiences. Hosted Agents give developers the control needed for complex, integrated, production-grade workloads.
For Australian businesses, the right choice should be based on risk, data sensitivity, integration needs, operational maturity, and compliance expectations — not just technical preference.
Our team can help assess where AI agents fit into your Azure environment, identify the right architecture, and design a secure path from pilot to production.