Skip to main content

​
Zero-Config Observe Mode

observe() is the easiest way to start with TealTiger. One function call, no configuration, immediate visibility into your agent’s behavior.

​
Progressive Disclosure

TealTiger follows a four-level progressive disclosure path. You start at Level 0 and graduate when you’re ready: Progressive disclosure levels Each level is additive. Level 1 includes everything from Level 0. Level 2 includes everything from Level 1. You never lose capabilities by graduating.

​
How observe() Works Internally

When you call observe(client), TealTiger creates a transparent proxy around your provider client: observe() proxy request flow

​
Key Properties

  1. Transparent: The proxy exposes the same API surface as the original client. Every method, every parameter, every response — identical.
  2. Non-blocking: Nothing is ever blocked in observe mode. PII detection logs findings but never interferes with the request.
  3. In-process: No network calls, no proxies, no sidecars, no external services. Everything runs in your process.
  4. Deterministic: No LLM in the governance path. Same input always produces the same instrumentation behavior.
  5. Low overhead: Less than 5ms added per request (P99). The provider network call dominates total latency.

​
Proxy Wrapping

The proxy uses language-native mechanisms to intercept method calls:
  • TypeScript: ES6 Proxy with handler traps for get and apply
  • Python: __getattr__ delegation with async/sync method detection
The proxy detects which methods trigger LLM calls (e.g., chat.completions.create, messages.create) and instruments only those. Utility methods, configuration setters, and non-API methods pass through with zero overhead.

​
Cost Accumulator

The cost accumulator extracts token usage from provider responses and computes cost using TealTiger’s built-in pricing database:
Costs are aggregated at three levels:
  • Per-request: Individual call cost
  • Per-session: Sum of all requests in this observe() session
  • Per-agent: Sum of all requests across all sessions for this agentId
When the provider doesn’t return token usage (rare, but possible), the cost is recorded as zero with a pricing_unavailable flag.

​
PII Scanner

The PII scanner uses TealTiger’s existing PIIDetectionGuardrail running in REPORT_ONLY mode:
  • Scans both inputs and outputs for every request
  • Detects: Email addresses, phone numbers, SSNs, credit card numbers
  • Never blocks: Even if PII is found, the request proceeds normally
  • Logs findings: PII type, count, and request ID written to audit log (never the PII values themselves)
  • Fails silently: If the scanner errors internally, the request continues as if no scan occurred

​
Baseline Construction

The behavioral baseline automatically builds a statistical profile of your agent’s “normal” behavior:
  1. Collects metrics for the first N requests (default: 100)
  2. Records: latency, input tokens, output tokens, cost, tool call count
  3. Once N requests are processed, computes p50, p95, p99 for each metric
  4. Marks the baseline as complete and writes a baseline_complete audit event
After completion, the baseline is frozen — subsequent requests don’t modify it. Access it via:

​
When to Use observe() vs Explicit Governance

The general rule: if you know what to block, use policies. If you don’t yet, use observe.

​
Supported Providers

observe() supports all 14 LLM provider clients: If an unsupported client is passed, observe() throws an UnsupportedProviderError identifying the type.

​
Air-Gapped and Offline Deployment

observe() works in air-gapped environments with no internet connectivity. It:
  • Makes zero outbound network calls for instrumentation
  • Does not require a proxy, sidecar, or external service
  • Does not phone home, check licenses, or download updates
  • Stores all data in-process (memory + local audit output)
The only network call is the one your application already makes to the LLM provider.