Experiment candidates

Candidates are the prompts or agents you want to compare in an experiment. When you add a candidate, Fetch Hive captures the source configuration at that moment. Later edits to the source prompt or agent do not change the candidate already added to the experiment.

​

Candidate types

​

How do I add a dashboard prompt?

Open the experiment. Click Add candidate. Choose Dashboard Prompt as the type. Select the prompt. Click Add candidate. Use this when you want to compare the current prompt editor state.

​

How do I add a deployed prompt?

Open the experiment. Click Add candidate. Choose Deployed Prompt as the type. Select the prompt. Select the deployed prompt. Select the variant. Select the version. Click Add candidate. Use this when you want repeatable testing against a specific production or staging deployment version.

​

How do I add an agent?

Open the experiment. Click Add candidate. Choose Agent as the type. Select the agent. Click Add candidate. Agent candidates run each dataset row as an isolated non-streaming call. They do not add messages to the source agent’s dashboard chat history.

​

Candidate settings

Candidate settings are read-only in experiments right now. Edit model, temperature, tools, instructions, and other settings in the prompt or agent editor before adding the candidate. Experiment-local overrides are planned for a future version.

​

Refreshing a candidate snapshot

Use Refresh snapshot when the source prompt or deployed prompt has changed and you want future runs to use the latest source configuration. Refreshing updates the existing candidate’s captured snapshot. It does not change historical run results. Use this when you want to keep the same candidate row in an experiment instead of archiving and re-adding it. For deployed prompts, refresh uses the same selected deployment, variant, and version reference. If you want to test a different deployed version, add a new candidate for that version. For agents, refresh captures the current agent configuration as a new snapshot for future runs. Historical run results remain unchanged.

​

Archived candidates

If you archive a candidate, it is removed from future experiment setup and run selections. Existing run results remain available so previous comparisons still make sense. Archiving requires confirmation. Older experiments may show unsupported legacy candidates. They remain visible for historical results, but new runs cannot include unsupported workflow snapshots. Archive unsupported candidates before starting a new run. See also: Prompts and Publishing and versioning