Experts & arenas

An arena is a supervised, multi-agent run: a coordinator agent drives a set of participant agents toward an objective, and the run ends when a success predicate is met. It is the platform’s mission-control view for agents working together (or against each other) on a goal.

This page documents the Arenas UI at /studio/arenas. For the conceptual model of experts and skills, see Skills & experts — this page is the complementary “where do I watch and steer a run” surface.

Arenas list

The list page shows one card per arena. The header summarizes counts (running / paused / won). Controls:

Filter arenas… text box — matches on name and objective.
Status filter chips: All, Running, Paused, Won, Failed, Terminated (each with a count).
+ CREATE ARENA — opens the arena wizard.

Each card shows the objective type, objective text, live status, elapsed time, participant count and avatars, a top-line metric with a sparkline, and progress toward the predicate. Cards flag N pending approvals and policy paused states. Click a card to open its detail page.

Objective types

Type	What it is
Collaborative	Participants work together toward a shared goal.
Adversarial	Participants compete; a judge-style coordinator awards rounds.
Dataset generation	Agents generate and curate accepted examples toward a target count.
Other	Open-ended objective.

Success predicates

An arena ends when its predicate is satisfied. Cards render the predicate as progress:

Predicate	Shown as
Metric threshold	`current / value` bar for a named metric.
Dataset target	`accepted / target` count bar.
Coordinator judged	The coordinator decides (no numeric bar).
Python snippet	A custom Python predicate (shows a snippet preview).

Create Arena wizard

+ CREATE ARENA opens a stepped wizard:

Type & name — pick the objective type and name the arena.
Objective — describe the goal and choose the predicate kind (metric threshold, coordinator-judged, or Python snippet).
Coordinator — choose a coordinator template. Options are gated by objective type, e.g. research-coordinator (collaborative/other), tournament-coordinator (adversarial), curation-coordinator (dataset generation), or a custom coordinator you author.
Participants & authorized set — choose the participant agents and the set of agents authorized to be activated. Initial participants are auto-added to the authorized set.
Policy — (dataset-generation arenas only) curation/acceptance policy.
Budget & caps — spend and resource limits.

Arena detail page

/studio/arenas/$arenaId is a three-pane mission-control view. A header strip shows the objective type, workspace repo, creator, and the full objective, with lifecycle actions gated by status:

RESUME (when paused), PAUSE (when running), TERMINATE (when running or paused).
IMPORT TO DATASET — shown for dataset-generation arenas, to pull accepted examples into a dataset.
BACK returns to the arenas list.

Participants (left pane)

Lists the coordinator plus every participant. Each entry is an activation of an authored agent (its own branch, model base, iteration, headline metric, heartbeat and lease). Select one to drill in; click to filter the timeline to just that agent.

Event timeline (center pane)

A live feed of arena events — commits, milestones, steering messages, promoted findings, blocks, approval requests, activations, accepted/rejected examples, policy changes, coverage gaps, and coordinator decisions. Can be filtered to a single participant (with a CLEAR FILTER button) or show “all participants · live”.

Mission / Agent (right pane)

Two toggle tabs:

Mission — overall mission state for the arena.
Agent: <name> — enabled once you select a participant; drills into that agent’s activity within the arena.

An Ask coordinator button sits in this pane’s header.