Build a Search as Code agent with the Perplexity API

The Agent API runs on prepaid credits, and every request needs a key from your API project.

1. 1Go to console.perplexity.ai and sign in with your Perplexity account.
2. 2Under Project, open Settings and fill in your project details if it asks.
3. 3In Billing, click Add payment method, then buy credits.
4. 4On the API Keys page, generate a key and copy it right away.

You should see a new key that starts with pplx-, shown on screen once.

Perplexity shows the full key only once. Save it in a password manager, and keep it out of code, prompts and shared files.

The Python SDK sends your requests and reads the key from your environment, so the key never sits in your code.

1. 1Open a terminal.
2. 2Run the commands below, with your key in place of [YOUR API KEY].
3. 3Check the install with pip show perplexityai.

mkdir sac-agent && cd sac-agent
python3 -m venv .venv
source .venv/bin/activate
pip install perplexityai
export PERPLEXITY_API_KEY="[YOUR API KEY]"

You should see the install finish with no errors, and a version number from pip show perplexityai.

The export line lasts only for this terminal window. In a new window, run it again before you run the script.

The model plans its searches from the full job, so it can split the work before it runs any code.

1. 1Save the script below as sac.py in the sac-agent folder.
2. 2Replace [YOUR TASK] with the whole job: the items, the facts you need and the format you want.
3. 3Keep both tools and max_steps as they are.

from perplexity import Perplexity

client = Perplexity()  # reads PERPLEXITY_API_KEY

TASK = "[YOUR TASK]"

response = client.responses.create(
    model="openai/gpt-5.6-sol",
    input=TASK,
    tools=[{"type": "sandbox"}, {"type": "web_search"}],
    max_steps=10,
    instructions=(
        "Write and run Python in the sandbox to do the searches. "
        "Plan every query first, run them in parallel, and drop duplicate URLs. "
        "Print only a short table of findings with a source URL for each. "
        "Then answer from that table."
    ),
)

for item in response.output:
    if item.type == "sandbox_results":
        print("--- Code the model ran ---")
        print(item.code)

print("--- Answer ---")
print(response.output_text)
print("Cost in USD:", response.usage.cost.total_cost)

You should see sac.py in your folder, with your own task in the TASK line.

A good task names every item and fact. For example: "Compare the free plans of 5 note-taking apps: storage, device limits and price, with a source for each."

The response returns the exact code the model ran, so you can see how it chose to search.

1. 1Run the command below and wait. A run with many searches takes a minute or more.
2. 2Read the printed code. Look for many queries sent at once, filters and duplicate checks.
3. 3Check the answer and the cost line at the end.

python sac.py

You should see one or more blocks of Python, then a short answer with sources, then the cost in US dollars.

With a direct model, max_steps starts at 1. The agent spends its first step loading the sandbox, so 1 step leaves no room to run code.

You change the plan in plain words, and the model writes new code on the next run.

1. 1Note what the first run missed, such as old pages, a skipped item or repeated sources.
2. 2Add a line like the one below to the end of instructions, inside the quotes.
3. 3Run python sac.py again and compare the new code with the first run.

Only use pages published after [DATE]. Search each item on its own, and prefer each maker's official site.

You should see new code with a date filter or one query per item, and a fresher answer.

To change speed and cost, swap the model for another one on Perplexity's Models page.