Foundry Hosted Agent Copilotkit

Developing with CopilotKit + AG-UI + Azure AI Foundry Hosted Agents

Use this skill for development work inside an EXISTING application built on this stack: a React/Next.js frontend using CopilotKit, connected over the AG-UI protocol to a Microsoft Agent Framework (MAF) agent (Python or .NET) that runs as — or is being developed against — an Azure AI Foundry hosted agent (paid Azure service; usage may incur costs).

Do NOT use this skill to scaffold a new project. Dedicated scaffolders exist (the CopilotKit CLI, azd ai agent init); use those, then return here for everything that follows: adding tools, gating them behind approvals, generative UI, shared state, debugging, dependency upgrades, and deploying agent updates.

Mental model

CopilotKit hooks (React)            useFrontendTool / useHumanInTheLoop /
        │                           useRenderToolCall / useCoAgent
        ▼
CopilotKit Runtime (route handler)  agents: { <name>: new HttpAgent({ url }) }
        │  AG-UI events over SSE
        ▼
AG-UI endpoint                      ← WHERE this lives defines your architecture
        │
        ▼
MAF Agent (tools, approval modes)   → model deployment

The single most important fact: a deployed Foundry hosted agent endpoint does not speak AG-UI by default. It exposes an OpenAI Responses endpoint (.../protocols/openai/responses) and/or a raw .../protocols/invocations endpoint. AG-UI must be produced somewhere, and where it is produced determines how every feature (especially human-in-the-loop) behaves. The three wirings are described in references/architecture.md (opens in a new tab).

Workflow

Follow these steps for every task on this stack:

1. Identify the wiring first. Inspect the codebase before changing anything:
   • add_agent_framework_fastapi_endpoint(...) (Python) or MapAGUI(...) (.NET) wrapping an in-process agent → Architecture A (in-process AG-UI endpoint).
   • A hosted agent whose own container serves AG-UI, declared with protocol: invocations in agent.yaml → Architecture B.
   • A separate service translating between the AG-UI endpoint and a hosted agent’s /responses endpoint (look for previous_response_id, mcp_approval_response, or a Foundry conversation object in the code) → Architecture C (translation bridge).
   • Confirm the frontend agent name: the key in the runtime agents config, the agent prop on the <CopilotKit> provider, and the hosted agent name in agent.yaml must all agree.
2. Ground in live documentation. Every layer here is pre-1.0 or preview and moves between minor versions. Never trust memorized APIs:
   • MAF and Foundry hosted agents: use the Microsoft Docs MCP tools when available, otherwise learn.microsoft.com (/agent-framework/integrations/ag-ui/, /azure/foundry/).
   • CopilotKit: docs.copilotkit.ai (Microsoft Agent Framework section). Verify hook and runtime API names against the TypeScript declarations bundled in the installed @copilotkit/* packages — names have churned (useCopilotAction is legacy; current names include useFrontendTool, useHumanInTheLoop, useRenderToolCall, useCoAgent).
   • AG-UI protocol: docs.ag-ui.com (event reference, dojo patterns).
3. Execute the task using the matching reference below.
4. Verify adversarially. A compiling build, a started dev server, or one successful chat reply is NOT proof. Apply the completion criteria at the end of this skill.

References

Load on demand; each is self-contained:

Task playbooks

Add or modify an agent tool

1. Define the tool on the agent (@tool in Python; AIFunctionFactory.Create in .NET) with typed, described parameters.
2. Keep docstrings grounding-safe: do not put concrete example values in parameter descriptions for fields the model must derive from real data — models copy literal examples. Use placeholders and validate inside the tool.
3. Return compact, model-consumable values; rich formatting belongs in the UI render, not the tool result.
4. Decide the approval mode now: side-effecting tools get approval_mode="always_require" (see references/hitl.md (opens in a new tab)); read-only tools stay unrestricted.
5. If the tool call should render in the UI, add a useRenderToolCall/render entry for it (references/patterns.md (opens in a new tab)).
6. Verify live: trigger the tool through the chat UI, confirm the call and result stream as TOOL_CALL_* events, and confirm renamed or re-typed parameters did not break any frontend component that parses the arguments.

Wire human-in-the-loop onto an existing tool

Follow references/hitl.md (opens in a new tab) end to end. Summary: mark the tool (approval_mode="always_require" / ApprovalRequiredAIFunction), enable confirmation on the AG-UI wrapper, register the approval UI hook on the frontend, and make the response payload shape match what the server detection expects. Then test approve AND reject AND a follow-up turn after approval (see the duplicate-execution hazard).

Build generative UI or shared state

Follow the pattern table in references/patterns.md (opens in a new tab). Know the honesty caveat: state synchronization patterns are native when the AG-UI adapter wraps an in-process agent (Architecture A/B); through a Responses-protocol bridge (Architecture C) they require explicit synthesis work — check what the codebase actually implements before promising the feature.

Debug a broken flow

1. Reproduce at the lowest layer first: curl -N the AG-UI endpoint with a minimal RunAgentInput JSON body and read the raw SSE events. If the bug reproduces there, the frontend is innocent.
2. For hosted agents, go one layer lower: call the agent’s /responses endpoint directly. This is how the known re-execution bug was isolated to the framework rather than the UI stack.
3. Match the symptom against references/troubleshooting.md (opens in a new tab) — exact error strings are listed.
4. Restart a locally running hosted agent (azd ai agent run) between verification passes if the agent holds in-memory state; stale state makes tests pass or fail for the wrong reason.

Upgrade dependencies

Follow references/upgrading.md (opens in a new tab). Never bump a single package in isolation: the version relationship rules there (runtime ↔ AG-UI client, agent-framework line consistency, hosting protocol ↔ manifest version) must hold simultaneously, and any local workaround must be re-validated against its tracked upstream issue before removal.

Deploy an agent update

Follow references/deploy-loop.md (opens in a new tab): iterate locally against the real agent with azd ai agent run, then azd deploy (each deploy creates a new agent version), then verify the deployed agent — including the approval pause — before declaring success.

Completion criteria

A change on this stack is done only when ALL of these hold:

1. The read/query path works through the real UI (not only via curl).
2. Every approval-gated tool was tested both ways: approve → the tool executes server-side and state visibly changes; reject → the tool does not run and the agent acknowledges.
3. At least one follow-up turn was sent in the same thread after an approval, and the gated tool did NOT silently execute again (references/hitl.md (opens in a new tab), duplicate-execution hazard).
4. Tool calls render correctly at stream end, not just during streaming (message snapshots can differ from live events).
5. For deployed changes: the checks above were run against the deployed endpoint, not only locally — deployment success is not proof of behavior.