Tools

Use these three tools before you build an agent. They run in your browser and send no data to any server. Each one keeps its state in the link: with the link, you share a result or a calculation.

Tool	What it is for	When
Agent, fixed workflow or person?	Answer ten questions about the task and get a recommendation with its reasons, risks and minimum controls.	Before you choose how to automate.
Compound reliability	See how often a chain of steps goes right and how many steps it can take. Compares pass^k (all attempts go right) with pass@k (at least one goes right) on sample runs.	When you design the flow and set the exit criterion.
Cost per successful task	Add up the model, the retries and human review, and spread it over the tasks that go right.	When you compare options, and every time prices change.

All three use the code that the repository tests. The evaluator comes from src/lib/decide and the calculators from src/lib/metrics. To put the decision in writing, go on to the templates.