Optional Module: Agentic Workflows

Outcome: A fixed loan-approval flow calls four MCP tools in a defined order.

Use a workflow when routing and parameter mapping should be configured in advance. Use an agent when the model must decide dynamically what to do next. Actual latency and cost depend on the tools and models used.

What we will implement

You will build four mock loan-processing tools, connect them in a deterministic @flow workflow, and expose that workflow through a native loan-processor agent.

1. Ask Bob to build the workflow

Ask Bob:

Build an ADK 2.14.0 loan-processing example with four mock MCP tools: credit
check, income verification, debt-to-income calculation, and final decision.
Create a deterministic @flow workflow that calls them in that order with
explicit map_input mappings, plus a native loan_processor_agent that uses only
the workflow tool. Put the files in tools/ and agents/, add requirements, run
local syntax checks, and validate all schemas. Do not import yet; explain every
input mapping.

Use these tested files as fallbacks or comparisons.

Save under tools/:

• loan_processing_server.py
• loan-processing-toolkit.yaml
• loan_approval_workflow.py
• requirements.txt

Download loan-processor-agent.yaml and save it under agents/.

The workflow performs:

credit check → income verification → debt-to-income calculation → decision

Review every map_input() call. It explicitly connects flow inputs and earlier tool outputs to the next tool's parameters.

2. Validate and import the MCP toolkit

Ask Bob:

Inspect the generated loan MCP server, toolkit YAML, workflow, and
requirements. Run Python syntax checks and any local smoke tests. If they pass,
import the MCP toolkit into draft and verify that all loan-processing-prefixed
tools are listed.

Manual syntax and import fallback:

python -m py_compile \
  tools/loan_processing_server.py \
  tools/loan_approval_workflow.py

orchestrate toolkits import -f tools/loan-processing-toolkit.yaml
orchestrate tools list

Confirm that the loan-processing: tools appear.

The toolkit's command: python3 runs inside the Orchestrate runtime, not in the participant's local Windows or macOS shell.

3. Import the flow

Ask Bob:

Validate every required workflow argument and map_input expression against the
imported tool schemas. If they match, import loan_approval_workflow as a flow
tool and verify it is listed. Show me any inferred mapping before changing it.

Manual fallback:

orchestrate tools import \
  -k flow \
  -f tools/loan_approval_workflow.py

orchestrate tools list

Confirm that loan_approval_workflow appears.

4. Import and test the agent

Ask Bob:

Show me examples of loan requests I can send to loan_processor_agent to test
valid approvals, missing inputs, adverse decisions, and boundary values.

Manual fallback:

orchestrate agents import -f agents/loan-processor-agent.yaml
orchestrate chat ask --agent-name loan_processor_agent

Try:

Process a loan request for applicant APP-12345 for $250,000.

Check that the workflow runs once, required parameters are mapped, and the final decision explains the returned values without inventing them.

Checkpoint

• The MCP toolkit imports and exposes its tools
• Every required tool argument has a map_input() mapping
• The flow imports as a flow-type tool
• The agent calls loan_approval_workflow
• The response reflects the workflow result

Troubleshooting

A tool receives an empty parameter

Inspect the preceding node and the matching map_input() expression. Flow variables and tool outputs must be mapped explicitly.

The workflow tool is missing

Confirm that it was imported with -k flow, then inspect orchestrate tools list.

Additional challenges are available in Agentic Workflow Exercises.