M8 · Capstone & Hosting¶

Goal: combine everything — tools, harness, orchestration, evaluation, observability — into one small application. Then see how to ship it.

Self-paced. This module is designed to be completed after the workshop.

---

You've built every piece. The capstone wires a few of them together into a tiny trip-planning assistant, then points you at the hosting options for taking an agent to production.

1. Setup¶

In [ ]:

import sys, pathlib, warnings
sys.path.insert(0, str(pathlib.Path.cwd().parents[1]))
warnings.filterwarnings("ignore")
from workshop_utils import get_chat_client
from agent_framework import Agent, tool
from typing import Annotated
from pydantic import Field

import sys, pathlib, warnings sys.path.insert(0, str(pathlib.Path.cwd().parents[1])) warnings.filterwarnings("ignore") from workshop_utils import get_chat_client from agent_framework import Agent, tool from typing import Annotated from pydantic import Field

2. Capstone: a trip-planning assistant¶

We'll combine three things you learned:

• Tools (M2) — a weather tool and a currency tool,
• The harness (M4) — for built-in planning + memory + telemetry,
• Observability (M7) — usage middleware to watch cost.

A real build would add evaluation (M6) in CI and possibly multi-agent orchestration (M5). Start simple; grow as needed.

In [ ]:

from random import randint

@tool(approval_mode="never_require")
def get_weather(location: Annotated[str, Field(description="City name.")]) -> str:
    """Get the weather for a location."""
    conditions = ["sunny", "cloudy", "rainy", "clear"]
    return f"{location}: {conditions[randint(0,3)]}, high {randint(12,30)}°C."

@tool(approval_mode="never_require")
def convert_currency(
    amount: Annotated[float, Field(description="Amount in USD.")],
    to_currency: Annotated[str, Field(description="ISO code, e.g. JPY.")],
) -> str:
    """Convert USD to another currency (demo rates)."""
    rates = {"EUR": 0.92, "JPY": 157.0, "GBP": 0.79, "INR": 83.2}
    r = rates.get(to_currency.upper())
    return f"${amount:.0f} ≈ {amount*r:.0f} {to_currency.upper()}" if r else f"No rate for {to_currency}."

print("tools ready")

from random import randint @tool(approval_mode="never_require") def get_weather(location: Annotated[str, Field(description="City name.")]) -> str: """Get the weather for a location.""" conditions = ["sunny", "cloudy", "rainy", "clear"] return f"{location}: {conditions[randint(0,3)]}, high {randint(12,30)}°C." @tool(approval_mode="never_require") def convert_currency( amount: Annotated[float, Field(description="Amount in USD.")], to_currency: Annotated[str, Field(description="ISO code, e.g. JPY.")], ) -> str: """Convert USD to another currency (demo rates).""" rates = {"EUR": 0.92, "JPY": 157.0, "GBP": 0.79, "INR": 83.2} r = rates.get(to_currency.upper()) return f"${amount:.0f} ≈ {amount*r:.0f} {to_currency.upper()}" if r else f"No rate for {to_currency}." print("tools ready")

In [ ]:

from agent_framework import create_harness_agent
from collections.abc import Awaitable, Callable
from agent_framework import chat_middleware, ChatContext, ChatResponse

@chat_middleware
async def usage(context: ChatContext, call_next: Callable[[], Awaitable[None]]) -> None:
    await call_next()
    resp = context.result
    if isinstance(resp, ChatResponse) and resp.usage_details:
        print(f"   [usage] {resp.usage_details}")

trip_agent = create_harness_agent(
    client=get_chat_client(),
    max_context_window_tokens=128_000,
    max_output_tokens=8_192,
    name="TripPlanner",
    description="Plans trips using weather and currency tools.",
    agent_instructions=(
        "You are a friendly trip planner. Use the weather and currency tools when relevant. "
        "Produce a short day-by-day plan with practical tips."
    ),
    tools=[get_weather, convert_currency],
    disable_web_search=True,   # keep the capstone provider-agnostic
    middleware=[usage],
)

session = trip_agent.create_session()
query = "Plan a 2-day Tokyo trip. Include weather and tell me what $200 is in yen."

print("Assistant: ", end="", flush=True)
async for update in trip_agent.run(query, session=session, stream=True):
    if update.text:
        print(update.text, end="", flush=True)
print()

from agent_framework import create_harness_agent from collections.abc import Awaitable, Callable from agent_framework import chat_middleware, ChatContext, ChatResponse @chat_middleware async def usage(context: ChatContext, call_next: Callable[[], Awaitable[None]]) -> None: await call_next() resp = context.result if isinstance(resp, ChatResponse) and resp.usage_details: print(f" [usage] {resp.usage_details}") trip_agent = create_harness_agent( client=get_chat_client(), max_context_window_tokens=128_000, max_output_tokens=8_192, name="TripPlanner", description="Plans trips using weather and currency tools.", agent_instructions=( "You are a friendly trip planner. Use the weather and currency tools when relevant. " "Produce a short day-by-day plan with practical tips." ), tools=[get_weather, convert_currency], disable_web_search=True, # keep the capstone provider-agnostic middleware=[usage], ) session = trip_agent.create_session() query = "Plan a 2-day Tokyo trip. Include weather and tell me what $200 is in yen." print("Assistant: ", end="", flush=True) async for update in trip_agent.run(query, session=session, stream=True): if update.text: print(update.text, end="", flush=True) print()

!!! success "You combined four modules in ~30 lines" Tools (M2) + harness (M4) + observability (M7), with the harness quietly doing planning and memory (M3) for you. That's the payoff of understanding each piece on its own first.

3. Add a quality gate (M6)¶

Before shipping, wrap the capstone in an evaluation so regressions fail CI:

from agent_framework import evaluate_agent, LocalEvaluator, keyword_check

checks = LocalEvaluator(keyword_check("Tokyo"), keyword_check("yen"))
results = await evaluate_agent(agent=trip_agent, queries=[query], evaluators=checks)
results[0].raise_for_status()   # break the build on regression

4. Hosting: take it to production¶

An agent is just Python — you can host it many ways. Agent Framework provides first-class paths:

Option	Best for	Upstream sample
A2A protocol	exposing an agent as a standard endpoint other agents/frameworks can call	04-hosting/a2a/
Azure Functions	serverless, event-driven, scale-to-zero	04-hosting/azure_functions/
Durable Task	long-running, checkpointed orchestrations	04-hosting/durabletask/
Container	full control (Container Apps, AKS, anywhere)	04-hosting/container/
DevUI	a local chat UI to demo/debug an agent	agent-framework-devui

Expose an agent over A2A (sketch)¶

# Agent Framework can serve an agent as an A2A endpoint; another process wraps a
# remote A2A endpoint as an `A2AAgent` and calls it like a local agent.
# See 04-hosting/a2a/ for runnable client + server.

The key idea: the agent code you wrote today doesn't change — hosting wraps it.

5. Where to go next¶

• Deepen orchestration: handoff, group-chat, and magentic patterns (03-workflows/orchestrations/) and checkpointed, human-in-the-loop workflows.
• Real memory: mem0, Redis, Azure AI Search, or the harness MemoryStore for durable cross-session memory.
• Model-graded evaluation: groundedness/relevance/coherence (05-end-to-end/evaluation/).
• Provider features: code interpreter, file search, hosted MCP tools (02-agents/providers/).
• MCP: expose your services as tools, or consume remote MCP servers (02-agents/mcp/).

📚 Agent Framework docs · microsoft/agent-framework

🧪 Capstone challenge¶

Pick one and build it:

1. Multi-agent trip planner — split into a researcher (gathers options) and a planner (writes the itinerary) chained with WorkflowBuilder agents-as-nodes (M5).
2. Evaluated assistant — write 5 evaluators for your capstone and wire raise_for_status() into a tiny CI script.
3. Hosted agent — follow 04-hosting/a2a/ to expose your capstone over A2A and call it from a second notebook.

---

🎉 You've gone from a single LLM call to a complete, observable, hostable agent. That's the whole agent harness — and you understand every piece of it.

← Back to Concepts · Home