# Niitaka

> AI agent observability and control plane. Monitor, debug, and govern AI agents
> in production via session tracking, event logging, policy enforcement, and
> experiment management.

## Documentation

### Getting Started

- [Overview](https://niitaka.ai/api/docs-md): what is niitaka, introduction, install
- [Quickstart](https://niitaka.ai/api/docs-md/quickstart): install, pip install, configure
- [Configuration & Auth](https://niitaka.ai/api/docs-md/config-auth): dotenv, credentials, secret

### SDK Reference

- [Core Concepts](https://niitaka.ai/api/docs-md/concepts): session, event, policy
- [Providers](https://niitaka.ai/api/docs-md/providers): openai, anthropic, gemini
- [Runtime Config](https://niitaka.ai/api/docs-md/runtime-config): get runtime config, model, temperature
- [Environments](https://niitaka.ai/api/docs-md/environments)

### Observability

- [Sessions](https://niitaka.ai/api/docs-md/observability/sessions): session, start session, running
- [Agent Dashboard](https://niitaka.ai/api/docs-md/observability/agents): agent id, register, lifecycle
- [Signals & Alerts](https://niitaka.ai/api/docs-md/observability/signals): signal, alert, guardrail

### Control

- [Policies](https://niitaka.ai/api/docs-md/control/policies): cost limit, step limit, retry
- [Policy Evaluation](https://niitaka.ai/api/docs-md/control/policy-evaluation): evaluation, engine, decision trace
- [Analytics](https://niitaka.ai/api/docs-md/control/analytics): analytics, overview, cost

### Experimentation

- [What is an Experiment?](https://niitaka.ai/api/docs-md/experiments): experiment, a/b test, variant
- [Setting up variants](https://niitaka.ai/api/docs-md/experiments/variants): variant, baseline, traffic weight
- [Running experiments](https://niitaka.ai/api/docs-md/experiments/running): experiment id, report metrics, goal completed
- [Interpreting results](https://niitaka.ai/api/docs-md/experiments/results): results, forest plot, verdict
- [Limitations & constraints](https://niitaka.ai/api/docs-md/experiments/limitations): limits, max variants, 8 variants
- [Shipping a winner](https://niitaka.ai/api/docs-md/experiments/shipping): ship, promote, winner

### Guides

- [Guardrails & Policies](https://niitaka.ai/api/docs-md/guardrails): guardrail, cost limit, step limit
- [Cost Management](https://niitaka.ai/api/docs-md/cost-management): cost, budget, spend

## Key concepts

- Session: one complete agent run. Status is running|done|stale — never "error". Check error_count > 0 for failures. Optional `tags` dict for custom key-value filtering. start_session() returns nullcontext() when unconfigured (null-object mode). create_session() for fire-and-forget parent session pre-creation.
- Event: structured log entry (llm, tool, decision, error). Latency is in event.data.meta.latency (seconds).
- Policy: real-time rule evaluated on every event. Types: cost_limit, step_limit, retry, fallback.
- Runtime Config: unified config schema {llm: {model, temperature, max_tokens, system_prompt}, guardrails: {cost_limit_usd, max_steps}}.
- Signal: structured alert emitted by the policy engine. Types: guardrail, control, alert, policy.
- Experiment: randomised A/B comparison of agent variants with statistical analysis.

> Full content (all pages concatenated): https://niitaka.ai/llms-full.txt