List variations
Every variation on an agent, with the score that decides which one runs.
info arrives whether you ask or not
Unlike tool calls, where info is gated, this endpoint returns the full info block on every request. includeInfo changes nothing:
score, feedbackCount, assignments, memoryLayerAssignments, model, createdBy, and the four counts (toolCount, toolSetCount, subAgentCount, memoryLayerCount). Nothing to opt into.
score is a Beta posterior, not an average
score is the posterior mean of Beta(ts_alpha, ts_beta), on [0, 1], where 0.5 is neutral. A fresh variation with no feedback reads exactly 0.5, because the prior is Beta(1, 1).
Each feedback score, on [-1.0, 1.0], moves it. Watch one variation take three:
ts_alpha and the negative on ts_beta, so the score is α / (α + β).
That is why it is not an average. Three positives out of three reads 0.8, not 1.0: the prior keeps pulling toward neutral until enough evidence accumulates. A variation with one lucky win cannot outrank one with fifty solid ones, which is the entire point when weighted selection samples from these posteriors.
0 is recorded and increments feedbackCount, but moves neither ts_alpha nor ts_beta. The score stays where it was. Use it to log a run you have no opinion about, not to signal mediocrity.[-1.0, 1.0] is a 400.
Reading the leaderboard
sortOrder sorts by creation, not by score. Sort on the client when you want a ranking:
feedbackCount alongside score every time. A 0.5 with n=0 is an untested variation; a 0.5 with n=200 is a genuinely mediocre one. The number alone cannot tell them apart.
What each variation carries
labels to slice a large agent. limit sets a page size the SDK iterator pages past on its own.
Related
Submit feedback
[-1.0, 1.0] score that moves the posterior.Create an agent
variationSelectionMode, and what WEIGHTED samples from.Update a variation
updateMask.Add an assignment
info.assignments.Authorizations
Bearer authentication header of the form Bearer <token>, where <token> is your auth token.
Path Parameters
Workspace ID.
"workspace_01HXKD2E5NQM3T9AYWCF133E3Q"
Agent ID. Accepts the canonical agent_… form or the external_id:<value> form.
"agent_01HXKD2E5NQM3T9AYWCFMGWT9Y"
Query Parameters
Maximum number of results to return
Pagination cursor from previous response
Sort order for results (asc or desc by creation time)
When true, the info field on each returned variation is populated.
Requests with this flag count more against your rate limit.
Filters by metadata labels. Comma-separated key=value pairs, e.g. "env=prod,team=ai". A resource matches only if every pair matches exactly (AND semantics).