Simulated SandboxTool Execution Barrier

Autonomous agent replay: with vs without Gate

Simulated iOS screen recording replay. Watch how Gate deterministically evaluates policies, halts high-risk actions at the execution barrier, and cryptographically signs tamper evident audit receipts.

Simulated Demo — Interactive UI mockup & scripted screen recording for illustration purposes.

Sonnev Gate Replay Engine

Scenario:

Protected by Sonnev Gate Barrier

09:41

Claude Sonnet · Agent

Message Claude...

Execution Forensics

Deterministic Barrier

Strict `flights.search` perimeter. Raw PII reads blocked. £595 charge held for FaceID sign-off.

Enforced Guardrails:

Mailbox & cloud drive access hard-denied by Gate perimeter

Deterministic threshold barrier holds payments > £250

Cryptographically signed Ed25519 receipt generated upon sign-off

SHA-256 tamper-evident audit ledger entry recorded

Cryptographic Audit LedgerSHA-256 Chained

CURRENT LEDGER HEAD:

8f4e2b10a9c81736b4e05912d38561947bcf0a8162394e5781a93d8b1e4c7602

Read the Gate Perimeter Architecture

Scenarios

How Gate governs agent actions

1. Low risk actionsAuto allow

Actions below specified policy thresholds (e.g. read operations or disbursements ≤ $500) evaluate in sub-millisecond local latency and execute immediately without operator friction.

2. Critical disbursementsApproval Queue

When a tool call exceeds automated limits, the execution halts safely. The request enters the operator approval queue until signed off or rejected by an authorized human.

3. Prohibited callsStrict Deny

Prohibited actions (e.g. dropping database tables or unauthorized privilege escalation) are rejected fail closed. Downstream tools are never invoked.

Ready to guard your agent fleets?

Integrate `@sonnevhq/gate` in TypeScript, wrap MCP servers with `@sonnevhq/mcp-proxy`, or sign in to configure policy.

Read developer docs Get started