Role-based tool access control for smolagents agents.

Demonstrate declarative YAML policy enforcement with real LLM calls.

Demonstrate role-based tool access control per crew agent.

YAML-defined data-sharing policies block PII and sensitive data.

Demonstrate Ring-2 tool capability enforcement with real function calling.

Demonstrate YAML policy enforcement blocking PII and internal access.

Model safety gates — restrict model downloads and inference trust.

Demonstrate trust-based quality gates using TrustedCrew.

Demonstrate behavioral anomaly detection with real LLM calls.

Behavioral anomaly engine detects rapid-fire tool calls.

Show governance blocking a secrets-access prompt before it reaches the LLM.

Demonstrate behavioral anomaly detection with rogue agent quarantine.

Full 4-agent pipeline: Research → Analyze → Summarize → Publish.

Run the full crew pipeline: Research → Write → Edit → Publish.

Adversarial prompt injection attacks blocked by YAML policy.

Demonstrate governance blocking adversarial prompt injection attacks.

Unauthorized workflow delegation bypasses are caught.

Demonstrate governance enforcing proper workflow delegation chains.

Detect agents attempting tools outside their declared profile.

Demonstrate detection of agents attempting to use undeclared tools.

Merkle-chained audit trail with tamper detection.

Demonstrate Merkle proof generation and tamper detection.

Scenario A: Backend Telemetry Stream → Dashboard Widget Input: System monitoring detects latency spike Expected: Dashboard widget wi

Scenario 2: Abandonment Signal - Loss The agent starts helping but fails to engage the user effectively. User stops responding halfw

Scenario 3: Acceptance Signal - Success The agent provides a good response and the user immediately moves on to the next task withou

Run adversarial attack scenarios to test governance resilience. Probes 4 attack vectors: prompt injection, tool alias bypass, trust score man

Scenario B: User typing in IDE → Ghost Text Input: User types "def calculate_" in Python file Expected: Ghost text completion (not a

Scenario C: SQL Query Results → Interactive Table Input: User runs SQL query Expected: Sortable, filterable table (not text dump)

Scenario D: Time Series Data → Chart Input: Performance metrics over time Expected: Line chart visualization

Scenario E: Critical Error → Notification Input: System error with high urgency Expected: Toast notification (not buried in logs)

Scenario F: Automatic Modality Detection Shows how the engine automatically chooses the right modality based on context and data typ

Scenario G: React Component Code Generation Shows how the engine can generate actual React JSX code.

Scenario 1: Undo Signal - Critical Failure The agent provides a dangerous response and the user immediately reverts it (Ctrl+Z). Thi

Flag transactions involving FATF high-risk jurisdictions.

Screen a candidate against job requirements. Governance ensures fair, bias-free evaluation.

Screen name against OFAC SDN list (exact + fuzzy).

Redact sensitive content from a plain string.

Search plugins by name or description substring. Args: query: Case-insensitive search string. Returns: List

Wrap Microsoft Semantic Kernel with governance.

* Send notifications to approvers.

* Send audit event to enterprise webhook if configured

Send a message from one agent to another. Args: from_agent: Sender agent ID to_agent: Recipient agen

Send a task request to another agent.

Convert objects to JSON-serializable format.

Serialize the full message (header + ciphertext).

* Update the metrics endpoint. * * @param endpoint - New endpoint URL.

* Set policy enabled/disabled

Inject the AGT agent identity into Agent Shield. Args: agent_id: AGT agent identifier (DID or SPIFFE ID).

Set the policy check callback.

Set rollout status for an agent (canary, stable, rolling-back).

Set a custom similarity function for output comparison. Args: fn: Function taking (current_output, candidate_output) and

Set the conversation state. Args: state: The new conversation state

Set trust threshold for a specific function.

Set up a development environment

Set up a complete production environment

Map severity to Tailwind badge classes using VS Code theme vars.

* sha256(bytes) → lowercase hex. Used to deterministically derive the * canonical `did:mesh:<hex32>` server DID when a 409 conflict occurs * during

* Check if we should use mock response (local dev, no API key, etc.)

* Check if we should use mock response (local dev, no API key, etc.)