Research workflows
Which research tasks are a good fit for agents?
Decide whether dynamic decisions are needed before choosing an agent, workflow or script.
Start with task characteristics
Useful starting points are frequent tasks with clear inputs, outputs and a way to assess results. Literature organization, moving data between analysis tools and report preparation can anchor discovery, but feasibility depends on data, permissions and tools.
A fixed process does not need an agent
If steps and rules can be specified, a script or explicit workflow can be easier to inspect. Consider model-driven decisions when a task genuinely requires selecting tools or actions based on intermediate results.
Design human review into the workflow
A final “please verify” disclaimer is insufficient. Assign review responsibilities: sources in literature work, inputs and methods in analysis, evidence and conclusions in reporting. Unresolved uncertainty should pause execution or move the task to a person.
Evaluate representative tasks
Include normal inputs, missing information and failed tools. Compare total time, human corrections and traceability with the current process. One successful demonstration does not establish ongoing reliability.
Start with a focused project and keep run records before expanding.