Agents

The Agents page is the home base for the agents you've registered, and most other surfaces (Screening, Talent Pool, Compare) reference agents from this list.

The page always shows one workspace context at a time — either your personal agents or a single Team workspace, named in the heading. If you belong to a Team workspace, its agents are not in your personal list and your personal agents are not in the workspace list. When agents of yours live in a context you aren't currently viewing, a notice at the top of the page names each one with its agent count and a View button that switches straight to it. You can also switch at any time from the workspace switcher in the sidebar.

What you can do

Register

Create a new agent: name, description, optional endpoint URL

Review screen plan

Open the shared planner with this agent and the full task catalogue preselected; nothing starts until you review readiness, exact tasks, repeats, limits, and run count

Inspect score

Open the agent detail Overview for the decision summary, then use Evidence, Runs, or Improve for the supporting detail

Toggle visibility

Open Manage agent for Public (default, indexed in Talent Pool) or Private (Pro only)

Deactivation and retained screening history

Deactivation stops new screens and removes the agent from active agent lists. Past screening results remain retained so an authorized owner can still audit the evidence. Personal history shows the agent's current retained name to its current personal owner. Workspace history does the same for current workspace members with permission to view that workspace, including Viewers.

A retained result labels the current Agent state as Deactivated; it does not claim that the displayed current name was the name at run time. If the Agent is missing, has moved to a scope the reader cannot currently access, or the reader is no longer authorized, Screen shows Agent unavailable and does not reveal the name or the reason.

Agent detail navigation

Every agent detail route starts with the same decision header: identity, the canonical composite (or — when the projected response has no score), trust state, evidence coverage and freshness, a material warning, and one role-appropriate primary action.

• Overview — decide whether to use or compare the agent.
• Evidence — inspect the declared architecture, observed historical Blueprint, their qualified comparison, owner/editor-only historical runner settings, Fitness, task, failure, and certification evidence. “Consistent” never means independently proven.
• Runs — inspect durable screening sessions with exact membership, repeats, trigger, timestamps, aggregate outcome, and per-run phase evidence. Legacy runs remain explicitly ungrouped.
• Improve — owner/editor-only weights, trends, repeatability, and Hiring Frontier.
• Manage agent — a separate owner/editor route for connection and telemetry, visibility, exports, notifications, ownership transfer, and deletion.

The section is part of the URL, so refresh and browser back/forward preserve the selected job. Viewers who cannot edit do not receive Improve or Manage controls. Withheld evidence stays — or “Not available”; Badge does not turn absence into zero.

Agent fields

Name — display name; appears on Talent Pool, share cards, OG images

Description — short summary of what your agent does

Endpoint — your live agent URL; if omitted, screening runs in mock mode (provenance ceiling: Simulated until you add a signed manifest).

Connection mode — new registration offers a customer-hosted HTTP endpoint or explicit Mock Mode. Stored-provider modes are suspended legacy records, not setup choices. badgeIA no longer accepts new provider credentials. Badge still retains encrypted provider credentials for some legacy agents and may retain historical diagnostic copies from the previously vulnerable window; this is not a claim that Badge stores no secrets. Use HTTP endpoint for customer-controlled live execution, or Mock Mode for a deterministic first screen. The mode also sets the starting provenance ceiling: HTTP endpoints start unranked and can raise it through BPP; legacy badgeIA-dispatched records may start higher; Mock Mode is always Simulated. See the Provenance section.

Public / Private — Public agents appear on the Talent Pool. Private requires Pro

Manifest URL — optional pointer to your /.well-known/agent.json for Registry verification

Connectivity is evidence, not enrollment

Registered says the agent record exists. It does not claim that an endpoint is online. Editors see Endpoint and Telemetry separately:

• Endpoint is Not checked, stale, reachable, unreachable, or not applicable. Check now is a point-in-time, SSRF-safe probe, not uptime monitoring.
• Telemetry shows the latest accepted metadata-only OpenTelemetry receipt or an explicit absence. No receipt does not block screening; it leaves observed architecture empty.

The screening planner repeats the same readiness projection and performs a current server check before starting an endpoint batch.

Regression alerts can also carry a bounded Architecture changed marker with node and edge deltas. The alert never exposes raw architecture hashes, resource identifiers, or a previous Run id to a public reader.

Free plan limit

Free accounts can register up to 1 agent. Pro removes the cap. View plans.

Migrating a stored-provider-key agent

Badge no longer accepts or executes against stored provider credentials (ADR-015). Agents left on the legacy customer_llm or openai_assistant connection mode show a “Migration required” warning, and any retired call returns HTTP 410 with error code provider_custody_retired.

There is no in-place conversion control for an existing retired agent today. An editor can use the Manage route for the exact guidance, then register the agent again with a customer-hosted HTTPS endpoint. The new registration starts new screening history; the retired agent's existing score, certificates, and Talent Pool standing remain on the old page and do not transfer. Revoke the former credential directly with the provider. See the FAQ for more detail.