← All tools EXPERIMENT 01 / DECISION TOOL
Should this beagent? Start with the job, not the architecture. Six questions to help decide how much autonomy is useful.
~2 minutes No account or API key Runs in your browser
Try a real task shape
Daily research digest ↗ Investigate a failed build ↗ Issue customer refunds ↗
01 Does the task need AI judgement?Think about the work itself, before choosing a model.
Rules are enough Calculations, lookups or predictable transformations. It needs interpretation Messy language, ambiguous inputs or open questions.
02 Can you write the steps ahead of time?A workflow can use AI without letting AI choose the whole process.
Yes, a fixed sequence The same steps in the same order. Mostly, with known branches The conditions and routes are predictable. No, results change the plan The next useful action depends on what it discovers.
03 What would it do outside the conversation?Reading a log and changing a customer account need different controls.
Return an answer No external tool actions. Read or search Retrieve information without changing it. Make changes Edit files, update records or take other actions.
04 How will you know it worked?A convincing explanation is not a check of the result.
A concrete check Tests, reconciled totals or an observable outcome. Someone can review it A person can judge the output before it matters. We have not defined that Success still means “looks about right”.
05 What happens when it gets something wrong?Consider the worst plausible mistake in this task.
A small inconvenience Cheap to spot, discard and retry. Work to undo it Changes are logged and can be reversed. A consequential mistake Money, access, people or hard-to-reverse changes.
06 Is this a repeated job?Automation has a maintenance cost as well as a running cost.
A one-off or early experiment We are still learning the task. It happens regularly There are enough examples to compare approaches.
YOUR TASK, YOUR TRADE-OFFS 0/6
Start with the task. Answer six questions to see whether a script, an AI workflow, an assistant or an agent is a useful starting point.
There is no score to chase. Changing one answer can change the recommendation.
Reset answers A transparent, rule-based design aid. No model call, saved answers or hidden scoring. Your judgement still matters.
Enable JavaScript to use the decision tool. The approach: use rules for predictable tasks, workflows for known steps, and trial an agent only when discoveries change the plan and you can check the result.
How the recommendation works This tool applies an editorial decision tree. It first checks whether success is defined and whether the work repeats. It then considers rules, consequences, the shape of the process and external tools. It does not predict performance or calculate a probability.
“Agent” here means a model choosing the next tool action from the results so far. A workflow can still use powerful models while code controls the steps.
Read every rule in the source ↗