Strands Build 1 · Give the agent a job, then stop it guessing

[akash-coded]

Give the agent a job

You are prototyping a desk-side assistant for your team. Before it can do anything clever, it needs two things: a clear role, and the sense to reach for a tool when it does not actually know something. That is what you will shape here.

Work at whatever level fits you. Everyone should clear Level 1; go as far past it as you like.

Spin up (30 seconds)

Run once, in Colab or locally. Stay on Haiku 4.5 while experimenting - it is fast and cheap.

pip install -q strands-agents strands-agents-tools

• Colab: open the Secrets panel (key icon, left sidebar) and add AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY.
• Local: run aws configure once, or export the AWS_ variables in your shell.
• Region: set it to wherever you enabled Bedrock model access (for example us-west-2).

import os
os.environ["AWS_DEFAULT_REGION"] = "us-west-2"   # your enabled-model region

from strands import Agent, tool
from strands.models import BedrockModel
from botocore.config import Config

def get_model(model_id="us.anthropic.claude-haiku-4-5-20251001-v1:0",
              temperature=0.3, max_tokens=2048):
    return BedrockModel(model_id=model_id, temperature=temperature, max_tokens=max_tokens,
        boto_client_config=Config(read_timeout=900, connect_timeout=900,
                                  retries={"max_attempts": 4, "mode": "adaptive"}))

def say(result):
    return "".join(b.get("text", "") for b in result.message["content"]).strip()

MODEL = get_model()
print("ready:", MODEL.get_config()["model_id"])

If something complains:

Message	Fix
on-demand throughput isn't supported	keep the us. prefix on the model id
throttling / too many requests	you are fine - the retries handle it, wait a few seconds
could not load credentials	recheck the region and keys above

Level 1 - give it a role

Create one agent with a personality and a job, then talk to it.

buddy = Agent(model=MODEL, system_prompt="...you decide the role...", callback_handler=None)
print(say(buddy("...your question...")))

• Pick a role you would actually use: a concise travel concierge, a patient Python tutor, a blunt code reviewer. Your call.
• Ask it three things. Then change one line, the system prompt, and ask the same three things again.

You are done when: you can show two clearly different "personalities" answering the same question, changed only by the system prompt.

Level 2 - stop it from guessing

Your agent will happily invent things it cannot know: today's date, a live price, an order status. Give it one tool so it looks things up instead.

@tool
def order_status(order_id: str) -> str:
    """Look up the delivery status of an order.

    Args:
        order_id: the order number, e.g. "1042"
    """
    # return something, real or faked - the wiring is the point
    ...

agent = Agent(model=MODEL, tools=[order_status], callback_handler=None)
print(say(agent("Where is order 1042?")))

• The docstring is not decoration. The model reads it to decide when to call the tool. Write it like an instruction.

You are done when: the agent answers an order question by calling your tool, not by guessing.

Level 3 - push it

Pick any one:

• Inspect the work. Print result.metrics.get_summary() and find total cycles and total tokens.
• Turn the dial. Build the agent with temperature=0, ask a creative question twice; rebuild with temperature=0.9, ask twice. Compare.
• Add a second tool and ask a two-part question. Check agent.tool_names, then look in the metrics summary to see which tools fired.

You are done when: you can explain in one sentence each what temperature changed, and how the agent decided which tool to use.

Notice this

• The system prompt is the steering wheel. Small wording changes move behavior a lot.
• You never wrote routing logic. The model decides when a tool is relevant, from its name and docstring.
• Tokens and cycles are your cost and behavior signals. Watch them early.

Poke at it

Ask the agent something your tool does not cover. Does it still try the tool? Why might it, and how would you stop it?