Develop strategies and test them through computer use

Use the free Computer Use mission to improve a bounded procurement recommendation. Start with the guided example, compare constructed strategies, and inspect why a changed question can reveal a better plan without weakening the buyer's requirements. No purchase is made.

The browser laboratory works without an account, wallet or API key. It prepares an inspectable recommendation and a task for a computer-using worker. Running a scripted browser exercise, running a model provider, checking an outcome and authorizing real work are separate steps.

Start with the buyer's actual need

The original fictional example needs 40 units in total, with at least 10 available by day 7 and all 40 by day 14. The budget is USD 510.00 (51,000 cents). Its three public demo offers are:

Buying all 40 from Stellar costs USD 490.00 and meets the need. Buying all 40 from Laurentian costs USD 300.00 but misses the early requirement. Buying all 40 from Boreal costs USD 560.00 and exceeds the budget.

A staged plan buys 10 from Stellar and 30 from Laurentian. Including both shipping charges, it costs USD 360.00 and meets both deadlines. The USD 130.00 difference is a calculation on this disclosed fictional scenario; it is not observed customer savings or a claim about real supplier availability.

A conventional two-delivery comparison also finds USD 360.00. The constructed strategy does not uniquely discover this plan or outperform that credible alternative. Its comparison with USD 490.00 changes the extra workflow assumption; it is not a same-task algorithmic improvement.

If the buyer instead requires all 40 by day 7, that becomes a hard requirement. The later Laurentian shipment cannot be used to make a cheaper plan appear acceptable. Stellar's USD 490.00 plan remains the feasible comparison. A useful reframing preserves the real need; it cannot rewrite an inconvenient deadline.

What Level 3 and Level 4 mean here

Level 3 develops the strategy itself. The bounded search constructs and revises typed programs for examining offers and planning quantities. Inspect the program, its parents, the changed operations, the development results and the failed attempts. The method is frozen before a fresh case is examined. Changing a program after that test creates a new candidate, rather than rewriting the old candidate's record.

Level 4 challenges an extra assumption. Buying from one supplier is one way to meet the buyer's demand, but it is not necessarily part of the demand. A staged allocation can change that formulation while preserving quantities, deadlines, budget and the requirement to make only a recommendation. The laboratory shows the original and reframed plans with their checks. Approval authorizes a local comparison only. It does not authorize procurement or establish general transformational intelligence.

The search space and checker are deliberately bounded. A local program search can demonstrate a different strategy on these cases; it does not demonstrate open-ended invention or autonomous business judgment. Visible controls and fresh cases help examine the result, but do not create protected independent proof.

Use the public procurement desk

Open the desk from the mission. Read the brief and supplier offers, enter a proposed allocation, and ask the desk to check it. Its quantity, deadline and budget checks give a concrete outcome. The same public desk can be operated with a keyboard and mouse by a person or an authorized computer-using worker.

Download the mission record before leaving. Keep the strategy, its lineage, the frozen comparison, the requirements and the recommendation together. Reopening a record checks its bindings; a checksum authenticates neither an author nor a reviewer. Imported evidence cannot grant authority or turn an unqualified result into independent proof.

The task export asks a worker to use the desk and retain the result. For the no-key route, copy the task into an existing ChatGPT Work session that has browser or computer access. Restrict it to the public demo desk and request screenshots and the final allocation. The task does not connect a worker automatically or guarantee that the selected session has the required capability.

Reproduce actual browser interaction locally

Start in the AGI Jobs checkout on Linux with /proc available. Use Python 3.12 for this setup; the retained Alpha runtime supports Python 3.11 through 3.13. Install the Node version in .nvmrc with your Node version manager. With nvm already installed:

nvm install
nvm use
npm ci
git clone --no-checkout https://github.com/MontrealAI/AGI-Alpha-Agent-v0.git ../AGI-Alpha-Agent-v0-computer
git -C ../AGI-Alpha-Agent-v0-computer checkout --detach 70dac8e5ce739978f5cd96d0fc190b6ebfcdf8d7
python3.12 -m venv .venv-successor-computer
. .venv-successor-computer/bin/activate
python -m pip install pydantic==2.14.0 httpx==0.28.1 cryptography==50.0.1 playwright==1.63.0
python -m playwright install --with-deps chromium
python integrations/alpha-agent/successor-computer-mission-lab.py \
  --alpha-source ../AGI-Alpha-Agent-v0-computer \
  --node "$(command -v node)" \
  --output reports/my-computer-positive

The output directory must be new. Playwright uses the browser installed for its exact package version. If your environment requires a reviewed local binary, add --chromium-executable /absolute/path/to/chromium; the evidence records that actual executable's identity. This host laboratory checks the trusted Alpha runtime closure and retains screenshots, action records, the task, the desk source and the checked allocation. Docker isolation requires its separate operator qualification.

The default local exercise uses a scripted adviser. Its browser actions and screenshots are real; its advice is predetermined. A successful run checks the computer-action transport and this task's outcome. It does not establish a frontier model's ability to discover the plan or operate the desk independently. Usage and cost remain unknown unless the actual provider supplies attributable records.

Run the two negative cases and prepare a separate operator package:

python integrations/alpha-agent/successor-computer-mission-lab.py \
  --alpha-source ../AGI-Alpha-Agent-v0-computer \
  --node "$(command -v node)" --output reports/my-computer-late --inject-error late
python integrations/alpha-agent/successor-computer-mission-lab.py \
  --alpha-source ../AGI-Alpha-Agent-v0-computer \
  --node "$(command -v node)" --output reports/my-computer-forged --inject-error false-allocation
python integrations/alpha-agent/successor-computer-mission-lab.py \
  --alpha-source ../AGI-Alpha-Agent-v0-computer \
  --node "$(command -v node)" --output reports/my-computer-operator --prepare-only

The negative modes apply to the default fixture. For a changed mission, supply --report /path/to/successor-computer-report.json in a positive or prepare-only run. The browser desk supports its three named supplier identities; an unsupported supplier profile is rejected. Inspect evidence.json, review.json and receipt.json from an executed case. The prepared package includes alpha-computer-task.json, supplier-desk.html, frozen-mission-report.json, execution-manifest.json, alpha-operator-commands.json, alpha-operator-commands.txt and chatgpt-work-mission.txt.

For the public no-key route, inspect the generated Work task and verify that the desk displays the exact requirements and offers. A changed mission must use its matching desk rather than silently falling back to the default example.

Provider credentials stay in the local operator environment. Do not paste them into the browser mission, task, export, command arguments or committed files.

Connect a model only through the separate operator route

The official OpenAI Computer Use guide, checked on 18 October 2026, documents the Responses API computer tool with gpt-6.1-sol. A returned computer_call contains ordered actions; the application executes them and returns a computer_call_output containing the resulting screenshot. For gpt-6-astra, the guide recommends code execution for computer interaction, while retaining the native computer tool as a supported alternative.

Use the separately installed, reviewed alpha-agent CLI and inspect reports/my-computer-operator/alpha-operator-commands.txt. Choose an available model and an admitted Docker image identified by its immutable digest. Set the generated command's local image, model, operator home and new result-directory placeholders. Provision the provider key through the named environment variable using your secret manager. Then prepare the exact task:

alpha-agent successor-computer prepare reports/my-computer-operator/alpha-computer-task.json \
  --image "$SUCCESSOR_COMPUTER_IMAGE" --model "$SUCCESSOR_COMPUTER_MODEL" \
  --provider responses --key-env OPENAI_API_KEY

Inspect the returned commitment. Place that exact value in SUCCESSOR_COMPUTER_PERMIT only when you authorize the bounded run, then use the generated command:

alpha-agent --home "$SUCCESSOR_OPERATOR_HOME" successor-computer run \
  reports/my-computer-operator/alpha-computer-task.json \
  --image "$SUCCESSOR_COMPUTER_IMAGE" --model "$SUCCESSOR_COMPUTER_MODEL" \
  --provider responses --key-env OPENAI_API_KEY \
  --permit "$SUCCESSOR_COMPUTER_PERMIT" --output "$SUCCESSOR_NEW_RESULT_DIRECTORY"

Configuration, account access, provider compatibility and isolation must be qualified on the actual task. The scripted lab result does not qualify either model or establish that a live provider call succeeded. An image name, model name or imported execution flag is not evidence that the corresponding runtime ran.

Know which evidence you have

Alpha is the advantage worth pursuing. The Gym develops methods. Specialist ASI requires earned, mission-specific superiority on agreed fresh protected work. SUCCESSOR Ω preserves useful methods and memory after proof and explicit bounded authorization. This laboratory awards no Specialist ASI designation and creates no execution, purchasing, payment or blockchain authority.