Practical notes: Agents, Tools, and Skills for a Working Mini AI Assistant

This walkthrough rebuilds the path from raw materials to a working system for: Agents, Tools, and Skills for a Working Mini AI Assistant. The focus is operable steps, explicit checks, and code that you can drop into a repo without guessing intent. For the Overview stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

What are tools, skills, and agents?

When working through the What are tools skills stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

get_current_conditions
get_forecast
get_alerts
get_historical_weather
get_air_quality
get_marine_conditions
get_river_conditions
get_wildfire_info
search_location
check_service_status

Temperature = Celsius, Fahrenheit, Kelvin

Length / Distance = Inches, Feet, Yards, Miles, Millimeters, Centimeters, Meters, Kilometers

Mass / Weight = Ounces, Pounds, Stone, Grams, Kilograms, Metric Tons

Volume / Capacity = Fluid Ounces, Cups, Pints, Quarts, Gallons, Milliliters, Liters, Cubic Meters

Area =  Square Feet, Square Meters, Acres, Hectares

Speed =  Miles per Hour (mph), Kilometers per Hour (km/h), Knots, Meters per Second (m/s)

Time Zones =  UTC/GMT offsets, Daylight Saving Time (DST) transitions, Unix timestamps to human-readable dates

Storage =  Bytes, Kilobytes (KB), Megabytes (MB), Gigabytes (GB), Terabytes (TB)

Experiment with Sample Code.

When working through the Experiment with Sample Code stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

Step 1 — Import the needed libraries

When working through the Step 1 Import the stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours. When working through the Step 1 Import the stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

import re
import os
import getpass
import requests

Step 2 — Create a calculator tool

The Step 2 Create a stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

def calculator_tool(prompt: str) -> str:
    print("    [Tool 1: Calculator] scanning prompt for dollar amounts...")
    dollar_amounts = re.findall(r'\$\s?(\d+(?:\.\d{1,2})?)', prompt)

    if not dollar_amounts:
        result = "no dollar amounts found"
        print(f"    [Tool 1: Calculator] {result}")
        return result

    values = [float(a) for a in dollar_amounts]
    total = sum(values)
    breakdown = " + ".join(f"${v:g}" for v in values)
    result = f"{breakdown} = ${total:.2f}"
    print(f"    [Tool 1: Calculator] found {len(values)} amount(s) {values} -> {result}")
    return result

Step 3 — Create the unit converter tool

The Step 3 Create the stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

UNIT_ALIASES = {
    "kilometers": "km", "kilometer": "km", "km": "km",
    "mile": "miles", "miles": "miles",
    "m": "meters", "meter": "meters", "meters": "meters",
    "ft": "feet", "foot": "feet", "feet": "feet",
}

DISTANCE_TO_MILES = {"km": 0.621371, "miles": 1.0, "meters": 0.000621371, "feet": 0.000189394}

def unit_converter_tool(prompt: str) -> str:
    print("    [Tool 2: Unit Converter] scanning prompt for distance legs...")
    legs = re.findall(r'(\d+(?:\.\d+)?)\s*(kilometers?|km|miles?|meters?|feet|ft)\b',
                       prompt, re.IGNORECASE)

    if not legs:
        result = "no distances found"
        print(f"    [Tool 2: Unit Converter] {result}")
        return result

    total_miles = 0.0
    breakdown = []
    for value, unit in legs:
        value = float(value)
        unit_norm = UNIT_ALIASES.get(unit.lower(), unit.lower())
        total_miles += value * DISTANCE_TO_MILES.get(unit_norm, 1.0)
        breakdown.append(f"{value:g} {unit_norm}")
    result = f"{' + '.join(breakdown)} = {round(total_miles, 2)} miles total"
    print(f"    [Tool 2: Unit Converter] found {len(legs)} leg(s) -> {result}")
    return result

Step 4 — Create the summarizer skill

The Step 4 Create the stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve. The Step 4 Create the stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

def summarizer_skill(text: str, max_sentences: int = 2) -> str:
    print("    [Skill: Summarizer] scanning message for key sentence(s)...")
    sentences = [s for s in re.split(r'(?<=[.!?])\s+', text.strip()) if s]

    if len(sentences) <= max_sentences:
        print(f"    [Skill: Summarizer] only {len(sentences)} sentence(s) -- returning as-is")
        return text.strip()

    stopwords = {"the","a","an","is","are","was","were","in","on","at","to","of","and",
                 "or","for","it","this","that","i","you","he","she","they","we","really"}
    words = re.findall(r'\b\w+\b', text.lower())
    freq = {}
    for w in words:
        if w not in stopwords:
            freq[w] = freq.get(w, 0) + 1

    scored = []
    for idx, sentence in enumerate(sentences):
        s_words = re.findall(r'\b\w+\b', sentence.lower())
        score = sum(freq.get(w, 0) for w in s_words)
        scored.append((score, idx, sentence))

    top = sorted(scored, key=lambda x: x[0], reverse=True)[:max_sentences]
    top_in_order = sorted(top, key=lambda x: x[1])
    summary = " ".join(s for _, _, s in top_in_order)
    print(f"    [Skill: Summarizer] kept {len(top_in_order)} of {len(sentences)} sentence(s) -> {summary}")
    return summary

Step 5 — Connect to your LLM

For the Step 5 Connect to stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

GROQ_API_KEY = os.environ.get("GROQ_API_KEY") or getpass.getpass(
    "Enter your free Groq API key (from https://console.groq.com/keys), "
    "or press Enter to skip: "
)

GROQ_MODEL = "openai/gpt-oss-20b"
GROQ_ENDPOINT = "https://api.groq.com/openai/v1/chat/completions"

def call_llm(augmented_prompt: str,
             system_prompt: str = "You are a helpful, concise assistant.") -> str:
    if not GROQ_API_KEY:
        return ("[No LLM reply -- no Groq API key was provided. Get a free one at "
                 "https://console.groq.com/keys, then re-run the setup cell above.]\n"
                 f"Here is the augmented prompt that would have been sent:\n\"\"\"\n{augmented_prompt}\n\"\"\"")
    try:
        response = requests.post(
            GROQ_ENDPOINT,
            headers={
                "Content-Type": "application/json",
                "Authorization": f"Bearer {GROQ_API_KEY}",
            },
            json={
                "model": GROQ_MODEL,
                "messages": [
                    {"role": "system", "content": system_prompt},
                    {"role": "user", "content": augmented_prompt},
                ],
                "temperature": 0.7,
                "max_tokens": 400,
            },
            timeout=30,
        )
        response.raise_for_status()
        data = response.json()
        return data["choices"][0]["message"]["content"].strip()
    except requests.exceptions.RequestException as e:
        return f"[LLM request failed -- {e}]"
    except (KeyError, IndexError, ValueError):
        return "[LLM returned an unexpected response format.]"

print("LLM configured." if GROQ_API_KEY else "No key entered -- running in fallback mode.")

Step 6 — Build your agents

For the Step 6 Build your stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary.

def show_step(step_num: int, label: str, content: str) -> None:
    """Small helper so every agent prints its pipeline the same, readable way."""
    print(f"\n  STEP {step_num} - {label}:")
    for line in str(content).splitlines() or [""]:
        print(f"    {line}")


def trip_planner_agent(prompt: str) -> str:
    """Agent 1. Condition: 2+ dollar costs AND 2+ distances -- a multi-stop itinerary."""
    print("[Router] -> Agent 1: Trip Planner Agent activated (detected an itinerary: multiple costs + multiple distances)")
    show_step(1, "Original prompt", prompt)

    cost_result = calculator_tool(prompt)
    show_step(2, "Tool result (Calculator -- total cost)", cost_result)

    distance_result = unit_converter_tool(prompt)
    show_step(3, "Tool result (Unit Converter -- total distance)", distance_result)

    augmented_prompt = (
        f"The user asked: \"{prompt}\"\n\n"
        f"A calculator tool already computed the total cost: {cost_result}\n"
        f"A distance tool already computed the total distance traveled: {distance_result}\n\n"
        "Using those two verified totals (don't redo either calculation yourself), give the "
        "user a short, friendly trip summary that reports both totals clearly."
    )
    show_step(4, "Prompt has changed -- now the augmented prompt sent to the LLM", augmented_prompt)

    reply = call_llm(augmented_prompt)
    show_step(5, "Final response from Groq", reply)

    return f"🧳 Agent 1: Trip Planner Agent:\n{reply}"


def text_agent(prompt: str) -> str:
    """Agent 2. Condition: default for any prompt, OR chained after Agent 1
    when the itinerary prompt also has an extra narrative sentence."""
    print("[Router] -> Agent 2: Text Analysis Agent activated")
    show_step(1, "Original prompt", prompt)

    summary = summarizer_skill(prompt)
    show_step(2, "Skill result (Summarizer)", summary)

    augmented_prompt = (
        f"The user wrote: \"{prompt}\"\n\n"
        f"Automatic summary of their message: {summary}\n\n"
        "Write a short, thoughtful, natural-sounding reply to the user that responds "
        "to what they actually said, informed by (but not just repeating) this summary."
    )
    show_step(3, "Prompt has changed -- now the augmented prompt sent to the LLM", augmented_prompt)

    reply = call_llm(augmented_prompt)
    show_step(4, "Final response from Groq", reply)

    return f"📝 Agent 2: Text Analysis Agent:\n{reply}"

Step 7 — Route to the correct agent

For the Step 7 Route to stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary. For the Step 7 Route to stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

def looks_like_itinerary(prompt: str) -> bool:
    dollar_amounts = re.findall(r'\$\s?\d+(?:\.\d{1,2})?', prompt)
    distance_legs = re.findall(r'\d+(?:\.\d+)?\s*(?:kilometers?|km|miles?|meters?|feet|ft)\b',
                                prompt, re.IGNORECASE)
    return len(dollar_amounts) >= 2 and len(distance_legs) >= 2

def has_extra_narrative(prompt: str) -> bool:
    sentences = [s for s in re.split(r'(?<=[.!?])\s+', prompt.strip()) if s]
    extra = [
        s for s in sentences
        if "quot; not in s
        and not re.search(r'\b(?:miles?|km|kilometers?|feet|ft)\b', s, re.IGNORECASE)
        and not s.strip().endswith("?")
    ]
    return len(extra) >= 1

def route_prompt(prompt: str) -> str:
    if looks_like_itinerary(prompt):
        reply = trip_planner_agent(prompt)
        if has_extra_narrative(prompt):
            print("[Router] -> also routing to Agent 2: Text Analysis Agent (extra narrative sentence detected)")
            reply += "\n\n" + text_agent(prompt)
        return reply
    else:
        return text_agent(prompt)

Step 8 — Testing your results

When working through the Step 8 Testing your stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

prompt = input("Ask me anything: ")

print("=" * 60)
print(f"PROMPT: {prompt}")
print("=" * 60)
final_answer = route_prompt(prompt)
print(f"\n  >>> RETURNED: {final_answer}")
print("-" * 60 + "\n")

Understanding the logic.

When working through the Understanding the logic stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

Step 1 — Routing to the Agent

When working through the Step 1 Routing to stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours. When working through the Step 1 Routing to stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Step 2 — Trip planner agent calls its tools

The Step 2 Trip planner stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

Step 3 — Call the calculator tool

The Step 3 Call the stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

Step 4 — Call the unit converter tool

The Step 4 Call the stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve. The Step 4 Call the stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Step 5 — Create updated prompt

For the Step 5 Create updated stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Prefer structured outputs with schema validation over free-form prose when the next step is code or a tool call.

Step 6 — Send results to GROQ

For the Step 6 Send results stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary.

Step 7 — Use text_agent

For the Step 7 Use textagent stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Authenticate at the gateway and re-authorize at the data plane. A bearer token alone is not a tenancy boundary. For the Step 7 Use textagent stage, define the inputs, the owner of the step, and the exit criteria before changing code. Operators should be able to re-run the step from a known checkpoint without guessing hidden state. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Step 8 — Use summarize skill

When working through the Step 8 Use summarize stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Keep configuration outside application code. Environment files, secret stores, and feature flags belong in one place operators can audit without reading the whole graph. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

Step 9 — Send results to GROQ again

When working through the Step 9 Send results stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours.

Making the case for agents, skills, and tools.

When working through the Making the case for stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Prefer small, testable units over sprawling scripts. When a step fails, the failure should point at a single responsibility rather than a tangled pipeline. Log tool name, args hash, latency, and outcome for every call. Debugging agent loops without that trail wastes hours. When working through the Making the case for stage, write down the contract first: required inputs, success signal, and what happens on partial failure. That checklist keeps later code changes honest. Record timings and token or query cost next to functional results. Cost visibility early prevents surprise bills when the path moves from demo to shared environments.

Operational checklist

The Operational checklist stage works best when treated as a measurable surface. Capture one golden transcript, one failure case, and the rollback note before expanding scope.

Treat this stage as a contract between inputs and validated outputs. Name the artifacts, define success checks, and refuse silent partial completion.

Expose tools with narrow schemas and explicit side-effect labels. Hosts need to know which calls mutate state before they auto-approve.

Put human approval on edges that spend money or change production data. Compile-time wiring does not equal business completeness.

Write a short runbook: how to rotate keys, how to drain the queue, how to roll back the last ingest.

Document the happy path and the recovery path together. Retries, human gates, and dead-letter handling are part of the product, not later polish.

Before promoting the stack, freeze versions, capture a golden transcript for the critical path, and confirm rollback steps. Shared environments need rate limits, tenancy checks, and a clear owner for secret rotation. Prefer boring reliability over clever one-off demos.

Batch note for 6caa2a29578e: keep provider keys out of the repo, set a per-session token ceiling, and store transcripts next to the eval fixtures so later model swaps stay comparable.