Observability & Auditing

Deploying AI agents to production requires deep, structured observability. Because agents can navigate workflows dynamically, standard HTTP logs are insufficient. Datafuse provides a comprehensive, session-aware Audit and Logging Pipeline that maps agent actions from start to finish.

Metric Tracking Capabilities

Datafuse intercepts every tool request at the gateway boundary, capturing core telemetry metrics:

• Execution Latency: Network-level breakdown showing proxy duration vs. provider response latency.
• Response Payload Audits: Raw byte sizes and response structures to alert on API contract shifts.
• Result Status Verification: Track downstream HTTP errors, database connection failures, and validation exceptions.
• LLM Session Tracking: Group multiple, sequential tool calls under a single parent agent session ID to trace complex multi-step reasoning workflows.

Dynamic Telemetry Visualizations

The Datafuse Console renders real-time execution statistics directly under /app/logs, grouping executions by latency curves, throughput, and error density.

    LATENCY DISTRIBUTION CURVE (Real-Time Console)

    500ms │               █
    400ms │             █ █ █
    300ms │           █ █ █ █ █
    200ms │         █ █ █ █ █ █ █
    100ms │       █ █ █ █ █ █ █ █ █
    0ms   └───────────────────────────────
             GET     POST    DELETE    PATCH

Structured Audit Schema

Every API invocation generates a strongly-typed audit trail.

Here is an example of an audit payload compiled by our telemetry services:

{
  "id": "e2f1e687-cc51-4f51-b847-ec074d0a68d0",
  "project_id": "proj_01j7v68s...",
  "session_id": "session_agent_alpha_92",
  "action": "slack.post_message",
  "resource_type": "toolkit",
  "resource_id": "slack",
  "duration_ms": 234,
  "status_code": 200,
  "metadata_json": {
    "ip": "12.34.56.78",
    "caller": "handshake_bot",
    "mutating": true
  },
  "created_at": "2026-05-23T03:07:43Z"
}

Decoupled Audit Logging

Developers can record custom application-level traces or operational messages directly through our unified SDKs.

This is highly useful for creating end-to-end debugging trails, linking LLM thoughts to external API calls:

from datafuse.generated import AuditLogCreate

# Record custom agent reasoning step in Datafuse console
sdk.logs.create_audit_log(
    audit_log_create=AuditLogCreate(
        action="agent.reasoning_step",
        resource_type="agent_reasoning",
        resource_id="session_123",
        metadata_json={
            "thought": "User requested channel publish. Searching for Ably integration...",
            "decision": "Invoke ably.publish_message"
        }
    )
)

Enforcing Zero-Retention Compliance

For industries governed by strict compliance laws (such as healthcare, banking, and government contracts), storing raw tool payloads on third-party servers presents a significant regulatory risk.

The Zero-Retention Vault

Datafuse offers a cryptographically enforced Zero-Retention Mode at the VPC container level:

• Ephemeral Proxy Ingress: Payloads are parsed completely in-memory inside container memory buffers.
• Instant Memory Purging: As soon as the payload response is returned to the client SDK, the buffer is zeroed out.
• Cryptographic Truncation: Logs retain only the metadata (timestamp, action slug, latency, status code) while the raw arguments and response structures are cryptographically redacted, ensuring zero PII leakages.

!WARNING When Zero-Retention Mode is activated, payloads cannot be recovered. Live debugging in the Console must be conducted in real-time via authenticated secure logging streams, as history data is never persisted.