Prompt file imported from microsoft/fsi-agentic-wealth-transfer (.github/prompts/new-foundry-agent.prompt.md). Fill in {{agent}} before use. Copyright stays with the author.

Scaffold a Foundry agent (Track B)

Name: new-foundry-agent
Author: microsoft

Build a real Foundry agent per foundry-agents.instructions.md and ../../docs/foundry-agent-architecture.md (SRD §5). The deployed Transfer Agent is reused, not rebuilt.

Input

Agent: {{agent}}. If omitted, ask which and its tools/grounding.

Steps

Read agents/transfer_agent/ (the existing hosted-agent pattern: main.py, runtime.py, _shared/audit.py, _shared/foundry.py) and azure.yaml (how the hosted agent is declared).
Realization:
- Risk → hosted agent (MAF, containerized, own Entra identity). Inputs ClientProfile + extractions; MCP tools for AML/OFAC/PEP (gated on Q2/Q3, start with a governed fake); output KYCResult{status: Approved|Review|Rejected, riskScore, reasonCodes, rationale, auditRef}; Review gates on a human; deterministic precedence/merge for conflicting sources.
- Exception → prompt agent (FoundryChatClient) that classifies the reject code into a route label only; the deterministic routing executors live in the workflow (/build-agent-workflow).
Governance: per-agent Entra identity; immutable WORM audit of every decision, fail closed; OTel gen_ai tracing; Content Safety/Prompt Shields; model via router (no hard-coded version); Foundry Evaluations set. Never auto-clear a human gate or fabricate a control number.
Ground the Exception classifier on Foundry IQ / Azure AI Search over the FINRA taxonomy (documents, not Fabric data).
Deploy/declare: add the hosted agent to azure.yaml/infra following the Transfer Agent; keep agent-framework-core.
Tests/evals: unit-test deterministic parts; add an eval set for the classifier; python -m pytest tests -q, ruff check agents ... tests, python -m evals.run_local.
Docs: update the RTM (/rtm-sync) and the agent architecture doc if realization changed.

Guardrails

Deterministic work is code/tools, not prompt logic. Reuse the Transfer Agent by reference.
Governed MCP tools behind APIM — no direct agent-to-endpoint calls. Human gates stay human-owned.

# Scaffold a Foundry agent (Track B) Build a real Foundry agent per [foundry-agents.instructions.md](../instructions/foundry-agents.instructions.md) and [../../docs/foundry-agent-architecture.md](../../docs/foundry-agent-architecture.md) (SRD §5). The deployed **Transfer Agent is reused, not rebuilt.** ## Input Agent: `{{agent}}`. If omitted, ask which and its tools/grounding. ## Steps 1. **Read** `agents/transfer_agent/` (the existing hosted-agent pattern: `main.py`, `runtime.py`, `_shared/audit.py`, `_shared/foundry.py`) and `azure.yaml` (how the hosted agent is declared). 2. **Realization:** - **Risk → hosted agent** (MAF, containerized, own Entra identity). Inputs `ClientProfile` + extractions; MCP tools for AML/OFAC/PEP (gated on Q2/Q3, start with a governed fake); output `KYCResult{status: Approved|Review|Rejected, riskScore, reasonCodes, rationale, auditRef}`; `Review` **gates on a human**; deterministic precedence/merge for conflicting sources. - **Exception → prompt agent** (`FoundryChatClient`) that classifies the reject code into a route label only; the deterministic routing executors live in the workflow (`/build-agent-workflow`). 3. **Governance:** per-agent Entra identity; immutable WORM audit of every decision, **fail closed**; OTel `gen_ai` tracing; Content Safety/Prompt Shields; model via router (no hard-coded version); Foundry Evaluations set. Never auto-clear a human gate or fabricate a control number. 4. **Ground** the Exception classifier on **Foundry IQ / Azure AI Search** over the FINRA taxonomy (documents, not Fabric data). 5. **Deploy/declare:** add the hosted agent to `azure.yaml`/infra following the Transfer Agent; keep `agent-framework-core`. 6. **Tests/evals:** unit-test deterministic parts; add an eval set for the classifier; `python -m pytest tests -q`, `ruff check agents ... tests`, `python -m evals.run_local`. 7. **Docs:** update the RTM (`/rtm-sync`) and the agent architecture doc if realization changed. ## Guardrails - Deterministic work is code/tools, not prompt logic. Reuse the Transfer Agent by reference. - Governed MCP tools behind APIM — no direct agent-to-endpoint calls. Human gates stay human-owned.

{ "prompt": "# Scaffold a Foundry agent (Track B)\n\nBuild a real Foundry agent per [foundry-agents.instructions.md](../instructions/foundry-agents.instructions.md)\nand [../../docs/foundry-agent-architecture.md](../../docs/foundry-agent-architecture.md) (SRD §5).\nThe deployed **Transfer Agent is reused, not rebuilt.**\n\n## Input\nAgent: `{{agent}}`. If omitted, ask which and its tools/grounding.\n\n## Steps\n1. **Read** `agents/transfer_agent/` (the existing hosted-agent pattern: `main.py`, `runtime.py`,\n `_shared/audit.py`, `_shared/foundry.py`) and `azure.yaml` (how the hosted agent is declared).\n2. **Realization:**\n - **Risk → hosted agent** (MAF, containerized, own Entra identity). Inputs `ClientProfile` +\n extractions; MCP tools for AML/OFAC/PEP (gated on Q2/Q3, start with a governed fake); output\n `KYCResult{status: Approved|Review|Rejected, riskScore, reasonCodes, rationale, auditRef}`;\n `Review` **gates on a human**; deterministic precedence/merge for conflicting sources.\n - **Exception → prompt agent** (`FoundryChatClient`) that classifies the reject code into a route\n label only; the deterministic routing executors live in the workflow (`/build-agent-workflow`).\n3. **Governance:** per-agent Entra identity; immutable WORM audit of every decision, **fail closed**;\n OTel `gen_ai` tracing; Content Safety/Prompt Shields; model via router (no hard-coded version);\n Foundry Evaluations set. Never auto-clear a human gate or fabricate a control number.\n4. **Ground** the Exception classifier on **Foundry IQ / Azure AI Search** over the FINRA taxonomy\n (documents, not Fabric data).\n5. **Deploy/declare:** add the hosted agent to `azure.yaml`/infra following the Transfer Agent; keep\n `agent-framework-core`.\n6. **Tests/evals:** unit-test deterministic parts; add an eval set for the classifier;\n `python -m pytest tests -q`, `ruff check agents ... tests`, `python -m evals.run_local`.\n7. **Docs:** update the RTM (`/rtm-sync`) and the agent architecture doc if realization changed.\n\n## Guardrails\n- Deterministic work is code/tools, not prompt logic. Reuse the Transfer Agent by reference.\n- Governed MCP tools behind APIM — no direct agent-to-endpoint calls. Human gates stay human-owned.", "variables": [ "agent" ], "notes": "Scaffold a Foundry agent (hosted or prompt) and wire it into the MAF workflow — Risk or Exception classifier. Mode: agent. Arguments: agent: risk | exception.", "metadata": { "source": { "provider": "github-prompts", "repository": "https://github.com/microsoft/fsi-agentic-wealth-transfer", "path": ".github/prompts/new-foundry-agent.prompt.md", "ref": "b02da046b94a1d13c0761b8e3e653fd540e38315", "url": "https://github.com/microsoft/fsi-agentic-wealth-transfer/blob/b02da046b94a1d13c0761b8e3e653fd540e38315/.github/prompts/new-foundry-agent.prompt.md", "key": "microsoft/fsi-agentic-wealth-transfer/.github/prompts/new-foundry-agent.prompt.md" } } }

new-foundry-agent

About

Scaffold a Foundry agent (Track B)

Input

Steps

Guardrails

Use it

Manifest

Reviews