Skip to content

Choose experiments for risk buy-down

Give URSA two small CSV files, ask an execution agent to rank candidate experiments under cost and schedule constraints, and then audit its selected campaign. This guided exercise emphasizes transparent calculations and reviewable artifacts rather than treating an agent's recommendation as a real qualification or safety decision.

You can run the same exercise in the TUI or dashboard. In either interface, make this directory the agent's workspace so the relative input paths resolve and the requested outputs stay beside the example.

Inspect the inputs

Open the design-risk table and the candidate-experiment table before launching an agent. The files use these schemas:

risk_id,risk_name,risk_area,initial_probability,impact,uncertainty,description

experiment_id,experiment_name,cost_kusd,duration_days,risk_targets,expected_uncertainty_reduction,confidence,description

The risk_targets field contains semicolon-separated risk IDs. All values and scoring rules are synthetic. Do not use the resulting ranking as evidence of readiness, qualification, or safety.

Prepare the example

Open a terminal in the URSA repository, enter this directory, and install its dependencies. The dashboard extra is included so either interface is available.

cd examples/single_agent_examples/execution_agent/risk_buydown
uv sync
export OPENAI_API_KEY="..."

Set-Location examples\single_agent_examples\execution_agent\risk_buydown
uv sync
$env:OPENAI_API_KEY = "..."

OpenAI works with URSA's built-in defaults and needs no config file. If you use another endpoint or model, configure it as described in the configuration guide.

Choose an interface

Start URSA from this directory:

uv run ursa

uv run ursa

Keep this example directory as the workspace. Paste the prompt in the next section with its leading #execute, which selects the execution agent. See the TUI guide for agent selection and application commands.

Start the dashboard from this directory:

uv run ursa-dashboard

uv run ursa-dashboard

Open the printed address, normally http://127.0.0.1:8080. Create an Execution Agent session and select this example directory as its workspace. Paste the prompt below without the leading #execute, because the session already selects the agent. See the dashboard guide for credential storage and workspace behavior.

Ask the agent to rank the experiments

Paste this prompt into your chosen interface. Keep #execute when using the TUI; remove only that prefix in an Execution Agent dashboard session.

#execute Read ./design_risks.csv and ./candidate_experiments.csv. Select an
experiment campaign that maximizes expected risk reduction subject to total
cost <= 150 kUSD, total duration <= 45 days, and at most 4 experiments.

For each risk use:
initial_risk_score = initial_probability * impact * uncertainty

For each experiment use:
expected_risk_reduction =
  sum(initial_risk_score for targeted risks)
  * expected_uncertainty_reduction
  * confidence

Search the feasible experiment combinations. Keep the calculation simple and
explainable. Write only under ./risk_buydown_outputs/ and create:
- experiment_rankings.csv
- selected_campaign.csv
- risk_before_after.csv
- experiment_value_scatter.png
- risk_before_after.png
- risk_buydown_reasoning.txt
- risk_buydown_recommendation.txt

Validate the input columns, do not overwrite inputs, and do not present this
toy analysis as a real qualification, readiness, or safety determination.

Review each proposed tool action before approving it. The execution agent can write files and run commands in its workspace; use a disposable copy when you adapt this exercise to data you cannot replace. The ExecutionAgent guide describes its tools and workspace behavior, and Sandboxing and information control explains stronger isolation options.

Audit the recommendation

Open risk_buydown_outputs/ after the agent finishes. Do not stop at the recommendation text. Check the work in this order:

  1. Confirm that both source CSVs are unchanged.
  2. Verify that experiment_rankings.csv contains every candidate and exposes the score components used for ranking.
  3. Recompute the selected campaign's total cost and duration from candidate_experiments.csv; confirm cost is at most 150 kUSD, duration is at most 45 days, and no more than four experiments were selected.
  4. Recompute the initial risk scores and expected reductions from the formulas in the prompt.
  5. Confirm that risk_before_after.csv accounts for every risk, including risks untouched by the selected campaign.
  6. Inspect both plots and make sure selected experiments are distinguishable from alternatives.
  7. Read risk_buydown_reasoning.txt and risk_buydown_recommendation.txt; require them to state remaining risks, assumptions, and the toy nature of the exercise.

If an artifact is missing or a constraint fails, ask the same execution-agent session to inspect its work and correct the output rather than silently accepting a partial result. Because an LLM chooses the implementation, exact rankings and filenames' internal formats may vary even though the requested files and constraints are fixed.