Implement agentic solutions
What makes it an agent
An agent = a model + instructions + tools + state, in a loop that decides which tool to call until the goal is met. In Azure AI Agent Service you define the agent, create a thread to hold conversation state, and start runs; the service executes tool calls server-side and appends results to the thread. You do not execute the tool loop yourself.
Tool selection
File search (vector store over uploaded documents), code interpreter (sandboxed Python for analysis and charts), Azure AI Search or Bing grounding (fresh or enterprise knowledge), Azure Functions / OpenAPI tools (business actions), and MCP servers (standardised external capabilities). Tool descriptions are prompts: vague descriptions cause wrong tool choice.
Orchestration patterns
Single agent with tools; connected agents where a manager delegates to specialists; or code-first frameworks such as Semantic Kernel, the Microsoft Agent Framework, or AutoGen. Prefer deterministic workflows for regulated or auditable steps, and reserve autonomous loops for genuinely open-ended tasks.
Traps: assuming code interpreter runs on your machine (it runs in a managed sandbox); assuming threads persist indefinitely (plan retention and summarisation); expecting memory to cross threads; forgetting tool schemas must be simple and that tool input/output consumes tokens.
Control and observability
Gate side-effecting tools behind human approval, cap loop iterations, scope credentials with managed identity and least privilege, and enable tracing so you can replay a run.
Rule of thumb: if you can draw the flowchart of the task, build a workflow and use the model only inside the steps - do not build an autonomous agent.