Header controls
The header is always visible regardless of which tab you are on.| Control | What it does |
|---|---|
| ← Back | Returns you to the Agent Foundry listing page. |
| Pause | Gracefully stops the agent. Any run currently in flight completes, and queued runs wait. Requires confirmation. |
| Resume | Re-activates a paused agent. Queued and scheduled runs begin processing again. Requires confirmation. |
| Restart | Immediately terminates any in-flight run and starts the agent fresh from the beginning. Requires confirmation. |
| Disable / Enable | Turns the agent off entirely (it stops firing and can’t be run) or back on. Shown as “Enable” once disabled; hidden when the agent has been suspended by an admin. Requires confirmation. |
| Status badge | Read-only indicator of the agent’s current state (Running, Evolving, Paused, Idle, Error). |
Restart terminates in-flight runs without waiting for them to finish. Use it to recover from a stuck or unresponsive agent — not as a routine reset. Always confirm the prompt before proceeding.
The eight tabs
- Overview
- Activity
- Memory
- Policy
- Data Context
- Outputs
- Evolutions
- Configuration
The Overview tab gives you a live at-a-glance picture of the agent’s health and current activity. All numbers update automatically in real time.Summary metrics:
Detail cards:
If the Overview snapshot fails to load, click Retry to reload it.
| Metric | Description |
|---|---|
| Total Runs | Cumulative run count, with a 7-day trend indicator. |
| Success Rate | Percentage of completed runs that succeeded, with a 7-day trend. |
| Avg Execution Time | Average duration of a completed run. |
| Confidence Score | Aggregated confidence across recent outputs. |
| Card | What it shows |
|---|---|
| Agent details | Name, description, owner, departments, and creation date. |
| Current task | The task the agent is working on right now, if any. |
| Task queue | Runs that are waiting to be picked up. |
| Resource usage | Live CPU, memory, and token consumption figures. |
The Activity tab is a live running log of every event the agent has produced, displayed newest-first.Filtering the log:
Click Clear filters to reset all active filters and return to the full log.Click any row to open a side panel showing the complete event detail, including the full payload, timestamps, and any associated metadata.
| Filter | Options |
|---|---|
| Kind | Trigger / Query / Insight / Alert / Action / Error / Retry |
| Since | Last hour / 24 hours / 7 days / 30 days / All time |
Use the Error kind filter combined with Last hour to quickly diagnose a recently failing agent without scrolling through unrelated events.
The Memory tab shows everything the agent currently remembers, and gives you tools to maintain and curate that memory over time.Memory type summary:
Each type shows an item count and a token count so you can gauge how much of the agent’s context window its memory occupies.Memory maintenance actions:
Filtering memory entries:
| Type | Description |
|---|---|
| Short-term | Working context from the current or most recent run. |
| Long-term | Retained knowledge promoted from earlier runs. |
| Episodic | Records of specific past events and interactions. |
| Semantic | General facts and conceptual knowledge the agent has learned. |
| Action | What it does |
|---|---|
| Run prune | Drops low-relevance entries to free up token space. |
| Run compact | De-duplicates and clusters similar entries into a smaller, denser set. |
| Promote | Moves selected high-value working-memory entries to long-term storage. |
- Source filter: All / Conversation / Document / Tool / Web
- Type filter: All / Short-term / Long-term / Episodic / Semantic / Procedural / Working / Periodic
| Bulk action | Notes |
|---|---|
| Delete | Permanently removes selected entries. There is no Trash window — deletion is irreversible. Requires confirmation. |
| Export | Downloads the selected entries as JSON or CSV. |
Bulk memory deletion is permanent. Deleted entries cannot be recovered. Review your selection carefully before confirming.
The Policy tab holds the rules this agent must always follow. These are the agent’s own pinned guardrails — distinct from the organisation-wide policies in Settings → Control Plane. Every save here records a new version of the agent, so there is always a trail of who changed a rule and when.
| Control / Button | What it does |
|---|---|
| Rules form | Set the agent’s standing rules from a structured form (the available rule fields are defined centrally, not typed free-hand). |
| Reset | Discards your unsaved rule edits back to the last saved state. |
| Save | Saves the rules and records a new agent version. |
| Custom instructions | Add or remove free-text instructions the agent should always follow, beyond the structured rules. |
| Suggestions from memory | Cards proposing rules the system noticed the agent tends to follow, ready to promote into permanent rules. Accept to add them. |
| History | A record of every policy change on this agent, for review and audit. |
The Data Context tab shows you what data the agent has access to and a full history of the queries it has run.Generated context card:Once the agent has produced a data context (a structured summary of the connected sources it has analysed), it appears here. Use this card to verify the agent’s understanding of your data model.Connected data sources:Each connected source appears as its own card showing:
- Source name and connection status
- A data quality bar indicating the overall health of the source
- The tables the agent can reach within that source
The Outputs tab collects everything the agent has produced: insights, alerts, and actions.Insights:Use the filter bar to narrow the insights list:
Click Acknowledge on any insight to mark it as reviewed. Acknowledged insights remain in the list but are visually distinguished from unreviewed ones.Alerts & Notifications:The Alerts & Notifications list is a delivery log showing every alert the agent has sent, the channel it used, and whether delivery succeeded.Actions Executed:The Actions Executed list shows actions the agent has proposed or carried out. For actions that require human approval, you can:
| Filter | Options |
|---|---|
| Severity | Low / Medium / High / Critical |
| Source kind | Anomaly / Trend / Forecast / Recommendation / Alert |
| Acknowledged only | Toggle to show only reviewed insights |
| Action | What it does |
|---|---|
| Approve | Authorises the pending action to proceed. Requires confirmation. |
| Reject | Declines the pending action. Requires confirmation. |
An action enters a pending state when its estimated cost exceeds the Approval threshold set in the agent’s Security configuration. You can adjust this threshold from the Configuration tab.
The Evolutions tab lets you improve the agent over time and review its full version history.Launching an improvement run:
The Ready badge appears next to the agent name when enough performance data has accumulated for a meaningful evolution run.Evolution timeline:The timeline shows every version of the agent in chronological order. The current version is expanded by default; earlier versions are collapsible. Click any version node to expand its details, including what changed and why.When an evolution run is in progress, a Progress panel opens automatically and shows live progress updates.
| Control | What it does |
|---|---|
| Auto-Evolve | Triggers a self-directed improvement run. The agent identifies its own weaknesses and adjusts. |
| Targeted Tune | Launches a focused improvement run aimed at a specific behaviour or failure mode you describe. |
| Major Upgrade | Initiates a larger-scope improvement run, potentially changing model configuration, prompts, or tool selection. |
Each evolution records a new version. You can review the diff between versions by expanding nodes in the timeline. Versions are not reversible through the UI — contact support if you need to roll back to an earlier state.
The Configuration tab is where you manage the agent’s settings after creation. Every change you save here is recorded as a new version in the version history.Version history card:Shows a digest of the three most recent versions. Click View all to jump to the full timeline on the Evolutions tab.Event triggers:Add, edit, test, and delete the agent’s triggers without going through the creation wizard.
Security context:Click Edit on the Security context card to update the agent’s role, permission level, PII handling, row limit, rate limit, and audit logging settings.Notification channels:Manage alert delivery bindings — add, edit, or remove channels without leaving the Configuration tab.Safety policy card:A read-only mirror of the current safety policy. Click Configure to open the security editor where you can make changes.Error log:Shows recent errors encountered by the agent. Use this alongside the Activity tab when diagnosing persistent failures.
| Action | How |
|---|---|
| Add trigger | Click Add trigger and choose Schedule, CDC, Webhook, or Manual. |
| Edit | Open the row menu (...) and select Edit. |
| Test fire | Open the row menu and select Test fire to run the trigger once immediately. |
| Delete | Open the row menu and select Delete. Requires confirmation. |
| Copy webhook URL | Available on Webhook-type triggers via the row menu. |
Permissions
Different roles have access to different controls in the agent detail view.| Role | Capabilities |
|---|---|
| Org Owner / Org Admin | Full access — pause, resume, restart, edit security and triggers, approve/reject actions, prune and delete memory, launch evolutions. |
| Dept Admin / Dept Manager | Operate agents within their department — monitor activity and outputs, acknowledge insights, approve actions, adjust configuration within department scope. |
| Employee | Typically view-only access to the Overview, Activity, Outputs, and Data Context tabs. Operational controls (pause, restart, memory delete, evolution runs) are reserved for higher roles. |
All destructive actions — Pause, Restart, bulk memory Delete, and trigger Delete — always prompt for confirmation before executing. Bulk memory deletion is permanent with no Trash window.