Release checks

Teams ship a web app several times a week, and the critical journeys such as sign in, search and checkout have to keep working. Classic browser tests break whenever a button moves or a class name changes, so somebody spends the sprint repairing selectors instead of writing tests.

A computer use agent looks at the rendered page the way a tester does. You describe the journey in words, and the agent finds the elements on screen.

The check

Point the agent at the staging URL and describe the journey.

from intelli.function.computer_agent import ComputerAgent
from intelli.function.browser_env import PlaywrightBrowserEnvironment

env = PlaywrightBrowserEnvironment(start_url="https://staging.example.com", headless=True)

agent = ComputerAgent(
    api_key=ANTHROPIC_KEY,
    provider="anthropic",
    model="claude-sonnet-5",
    environment=env,
    max_iterations=20,
    log=True,
)

try:
    result = agent.run(
        "Search for a laptop stand, open the first result, add it to the cart, "
        "then report the cart total and whether the checkout button is enabled."
    )
    print(result["output"])
finally:
    env.close()

Keep the agent read only

A smoke check should never place a real order. The on_action hook blocks any action you did not intend, and it blocks the action as well if the hook itself raises.

BLOCKED_WORDS = ["place order", "pay now", "confirm purchase"]

def guard(action):
    text = str(action.get("text", "")).lower()
    return not any(word in text for word in BLOCKED_WORDS)

agent = ComputerAgent(
    api_key=ANTHROPIC_KEY,
    provider="anthropic",
    environment=env,
    on_action=guard,
)

Run the checks in one flow

Each journey is a task, so the whole suite runs in a single flow and a final text agent turns the outputs into a release note.

import asyncio
from intelli.flow import Agent, Task, Flow, TextTaskInput

def journey_agent(mission):
    return Agent(
        agent_type="computer",
        provider="anthropic",
        mission=mission,
        model_params={
            "key": ANTHROPIC_KEY,
            "model": "claude-sonnet-5",
            "start_url": "https://staging.example.com",
            "max_iterations": 20,
        },
    )

sign_in = Task(
    TextTaskInput("Open the sign in page and report whether the form accepts a demo account."),
    journey_agent("check the sign in journey"),
    log=True,
)

search = Task(
    TextTaskInput("Search for a laptop stand and report how many results appear."),
    journey_agent("check the search journey"),
    log=True,
)

summary = Task(
    TextTaskInput("Write a short release check summary. Mark every failing journey clearly."),
    Agent(
        agent_type="text",
        provider="openai",
        mission="summarize release checks",
        model_params={"key": OPENAI_KEY, "model": "gpt-5.5"},
    ),
    log=True,
)

flow = Flow(
    tasks={"sign_in": sign_in, "search": search, "summary": summary},
    map_paths={"sign_in": ["summary"], "search": ["summary"], "summary": []},
    log=True,
)

results = asyncio.run(flow.start())
print(results["summary"]["output"])

Why it helps

The checks survive a redesign, because the agent reads the screen instead of a selector.
A new journey costs one sentence, so product people can add checks too.
The same flow runs against staging and production by changing start_url.

Release checks

The check​

Keep the agent read only​

Run the checks in one flow​

Why it helps​

The check

Keep the agent read only

Run the checks in one flow

Why it helps