AI Agent Center
The AI Agent Center is a live catalog of every AI agent in your project: agents sending traces, and agents listed from your Bedrock and Vertex AI accounts before they first run. It shows each agent's requests, cost, latency, risk score and status, so you can answer "what agents do we have, and is any of them misbehaving?" in one place. Each agent has a profile page with its full record, and a Suspend agent action stops it in one step.
Enterprise feature. Open it from Discover AI → AI Agent Center. It needs the AI Agent Center permission, which Owner, Department Admin, Security & Compliance and Cost Analyst have. In the custom role builder it is listed under Gauge AI. Without the Enterprise plan the page shows a preview with sample agents. See Members & Roles and Custom Roles.
How agents are discovered
There is nothing to register by hand. An agent is created in either of these ways:
- From traces. The first time a trace arrives with an agent ID and agent name, from the Python SDK, OpenTelemetry, a connector or the Portkey connector.
- From cloud inventory. The AWS Bedrock and Google Vertex AI connectors list every Bedrock agent and flow and every Vertex AI reasoning engine on a schedule and add them to the catalog, whether or not they have run. An agent listed this way that has never sent a trace is marked Not seen running, so dormant agents stand out.
Each agent is unique per project by name. Tags sent with its traces are merged onto the agent, and the most recent value of a tag wins.
Metrics come from the agent's traces for the selected time range. Risk signals from AI Inventory approvals, PII Detection and Drift Alerts are added on top. With the Last 24 hours window, the catalog refreshes every 5 minutes.
Headline tiles
The top of the page shows five tiles for the selected window:
| Tile | What it shows |
|---|---|
| Total Agents | Agents active in the window, "of N registered" |
| Total agent requests | Requests, with the number that succeeded |
| Total agent cost | Cost, with the token count |
| Avg agent latency | The mean of each agent's median (P50) latency |
| Avg agent risk | The mean of the agents' risk scores, as a live snapshot |
The agent catalog
Under Agent Catalog, switch between Grid view (cards, the default) and List view (rows). Agents are listed by last activity, newest first, 12 per page, with inactive agents at the end. The catalog shows up to 50 agents.
Each card shows the agent name and display name, a platform icon, its status badge, Requests, Cost, Latency (P50), Risk, the first two models (then "+N"), up to three tags, and Active with the last activity time or Inactive. If the agent used an unapproved model in the last 24 hours, the card also shows Last violation.
Each row in list view shows the same metrics, and a warning icon when the agent has issues. Its tooltip lists unapproved model traces, AI violations, drift alerts, error rate and slow P95 latency.
Platforms
The icon on each agent shows the platform it runs on, read from its tags: Bedrock Core and Vertex AI (from the AI Studio connectors), CrewAI, Claude Code, Codex, Copilot and Claude Cowork (from coding agent telemetry), and websites observed through the Zscaler integration. Agents with no platform tag get a generic avatar.
Status and risk score
The risk score runs from 0 to 100 and combines four parts:
| Part | Weight | What raises it |
|---|---|---|
| Security | 35% | AI policy detections (PII, PCI, PHI, ethics and bias, jailbreak) as a share of requests |
| Operational | 25% | Error rate, plus a penalty when P95 latency is above 5 seconds |
| Governance | 25% | Open drift alerts (five or more counts as the maximum) |
| Compliance | 15% | Unapproved model use as a share of requests |
Detection and unapproved model rates reach their maximum at 20% of requests. Open drift alerts count regardless of the time range. An agent with no requests and no signals has no score.
The status badge follows the score:
| Status | Risk score |
|---|---|
| Healthy | Below 10% |
| Warning | 10% to 49% |
| Critical | 50% or more |
| Inactive | No activity in the selected period |
A suspended agent also shows a Suspended badge on its card, its row and its detail panel. See Suspend an agent.
The detail panel's Risk tab also shows the four-dimension Risk Diamond Analysis used by the Risk Dashboard, which has its own scale (Low, Moderate, High, Critical).
Search and filters
- Search agents... matches display name, agent ID and model names.
- Time range: Last 24 hours (default), Last 2 days, Last 3 days, Last 7 days.
- Status: Healthy, Warning, Critical, Inactive.
- Model: any model the agents use.
- Cost: Low, Medium, High. The bands are set from your own agents' cost spread (lowest, middle and highest third) and recalculate when you change the time range.
- Cloud Type: Bedrock Core, Vertex AI, CrewAI and the other agent platforms.
- Filters builder:
- Compliance: Has Violations, No Violations (last 24 hours)
- Guardrail violations: PII, PCI, PHI, Ethics/Bias, Safety/Jailbreak
- Drift alerts: Has Active Alerts, No Active Alerts
- Integration: agents mapped to a Slack channel, Jira project or Linear team
- Tags: any tag key, with "any of" or "none of"
Clear All resets every filter.
Agent detail
Select an agent to open its detail panel. The header shows the agent name, its status and last activity, a violation badge for the last 24 hours (select it to jump to the Risk tab), and badges for the Slack channel, Jira project, Linear team, ServiceNow item and PagerDuty service linked to it. The panel has its own time range selector.
With the AI Agent Center permission at Read & Write you can rename an agent's display name with the pencil icon.
| Tab | What it shows |
|---|---|
| Overview | Key metrics (requests, success rate, cost, risk with its breakdown), models, tags, the guardrails configured for the project, and an activity chart |
| Performance | Response time at P50, P95 and P99, latency and error charts, and projected costs |
| Cost | Total cost and average cost per request, cost and token charts, and the at-scale estimate |
| Risk | Risk Diamond Analysis, AI violations by detection type with View Detection Traces, Unapproved Models, compliance status and Drift Alerts |
The Risk tab sections other than Unapproved Models also need governance read access (the Risk Dashboard and Agent Profiles & Drift permissions). Cost Analyst does not have it, so those sections do not load for that role.
Cost projections
| Card | How it is calculated |
|---|---|
| Your Projected Costs (Live Data) | The agent's average hourly cost in the window, scaled to a day, 30 days and 365 days |
| At-Scale Cost Estimate (Forecast) | The agent's average cost per request at 100,000 requests a day, scaled to a month and a year |
The live projection averages over hours in which the agent was active, so it is higher than actual spend for agents that only run part of the day.
Agent profile
Select Agent profile in an agent's detail panel to open its profile page. The page brings the agent's full record together in one place:
| Attribute | Source |
|---|---|
| Owner | Owning team from the agent's governance profile |
| Purpose | Business purpose from the governance profile |
| Deployment stage | From the governance profile |
| Risk tier | EU AI Act risk level from the governance profile |
| Data sources | Data sources declared in the governance profile |
| Models | Models the agent called, from its traces |
| Tools and MCP servers | Declared tools from the governance profile, next to the tools and MCP servers the agent actually reached, from AI TrailMaps |
| Permissions | The agent's Budget AI keys: model groups, allowed source networks, rate and spend limits, and expiry |
Tools the agent used that are not in its declared tools are highlighted, so unapproved tool use shows at a glance. Declared tools are recorded in the Declared tools field of the governance profile's evidence form.
Owner, purpose, stage, risk tier, data sources and declared tools are shown when the agent has a governance profile. An agent without one shows those attributes as not recorded, with a link to create the profile in Evidence Collection.
Export profiles on the AI Agent Center downloads a CSV with the profile of every agent in the project, one row per agent and one column per attribute.
Suspend an agent
Suspend agent in an agent's detail panel stops the agent in a single action:
- Every Budget AI key bound to the agent is blocked, in one transaction. All of them are blocked or none are. Gateways pick up the change within about a second and refuse the keys with 401
key_blocked. - The agent is marked suspended, with who suspended it and when, and shows a Suspended badge.
- Agents instrumented with the guarded clients of the ANTS Python SDK refuse model calls with an error stating that the agent is suspended. See AI Guardrails.
- When the Portkey connector is configured, the agent's Portkey keys are disabled through Portkey's admin API.
- The suspension is written to the audit log.
Resume agent puts the agent and the keys the suspension blocked back into service. It needs step-up verification. Suspending and resuming need the same permission as blocking and unblocking a key: gatewayAi.keys:admin or tenantSecurity.gatewayKeys:admin.
Suspending stops traffic that goes through Budget AI, the guarded SDK clients and, with the connector, Portkey. It cannot revoke credentials the agent holds directly in cloud or SaaS accounts, such as a provider API key or a cloud role. Revoke those in the provider or cloud account.
Unapproved models
Every model call from an agent is checked against your organization's approvals in AI Inventory. A model that is not approved, or not in the inventory at all, counts as a violation. Violations are kept for 7 days.
They appear as Last violation on the card, in the Has Violations filter, in the risk score, and in the Unapproved Models section of the Risk tab. That section shows how many traces used unapproved models, which models, when it was last detected, and View Traces. To clear it, approve the model in AI Inventory or change the agent.
Notifications
From the Unapproved Models section you can send a compliance notification to the Slack channel, Jira project or Linear team mapped to the agent. Drift alerts in the Risk tab can be sent to ServiceNow or PagerDuty. See Notifications for setup, mapping and message content.
Typical workflow
Scan for problems
Set the time range and filter Status to Critical or Warning, or use the Filters builder to find agents with violations, detections or drift alerts.
Drill into an agent
Open the agent and check the Risk tab to see which part is driving the score: detections, unapproved models, drift or errors.
Act
Approve or replace models in AI Inventory, review detections in AI Violations, work drift in Drift Alerts, and notify the owning team. If the agent must stop now, select Suspend agent.
Next steps
- AI TrailMaps: see what each agent called and where data went.
- Risk Dashboard: risk across all agents by dimension.
- Notifications: send findings to Slack, Jira, Linear, ServiceNow and PagerDuty.