Docs/Core Concepts/Observability

Observability Model

AI applications are non-deterministic. The same input can take a different path, call different tools and cost a different amount each time. ANTS Platform records each run in full so you can see what happened, what it cost and how good the result was.

ANTS Platform is built on traces. Metrics such as cost, latency, token usage and error rates are computed from traces, not collected separately.

Hierarchy

text
Organization └─ Project ├─ Agents └─ Traces (each with an environment, user, session, tags) └─ Observations (span, generation, event; nested) └─ Scores and comments

Organization

Your company account. Members, roles, SSO, billing, credits, data retention and organization-wide features are managed here under Organization Settings. See Organizations & Projects.

Project

The container for data. API keys, traces, dashboards, models, integrations and most settings belong to one project. Every SDK or exporter writes to the project that owns its key.

Environment

Every trace and observation has an environment, such as production or staging. The default is default. Set it with environment (or ANTS_PLATFORM_TRACING_ENVIRONMENT) in the SDKs, or deployment.environment in OpenTelemetry. Names must be lowercase letters, digits, hyphens or underscores, up to 40 characters. Every table and dashboard can be filtered by environment.

Agent

A named AI application, registered from the agent name your SDK sends or detected automatically for coding agents. See Understanding AI Agents.

Trace

One end-to-end run, such as one chat turn or one agent task. A trace has a name, timestamp, input, output, metadata, tags, user ID, session ID, environment, release and version. Latency, cost and token totals are computed from its observations.

Observation

A step inside a trace. Observations nest by parent to form a tree.

TypeUse it forExtra fields
SpanAny unit of work: a function, retrieval, tool call or agent stepStart and end time
GenerationAn LLM or embedding callModel, model parameters, token usage, cost, time to first token, linked prompt version
EventA point in time with no duration

Every observation has a level (DEBUG, DEFAULT, WARNING or ERROR) and can carry input, output, metadata and a status message.

Session and user

Traces with the same session ID form a session. Traces with the same user ID roll up into a user. Both are plain strings that you choose.

Score and comment

A score is a named value (numeric, categorical or boolean) on a trace, observation or session. Scores come from the API (source API), evaluators (EVAL) or human annotation (ANNOTATION). Comments are free-text notes on a trace, observation or session. See Scores.

Collecting data

MethodBest for
Python SDK, JS/TS SDK, Java SDKYour own applications and agents
OpenAI drop-in and LangChain callbackAutomatic tracing of LLM calls and chains
CrewAICrewAI crews, through the SDK listener or native OpenTelemetry
OpenTelemetry (OTLP)Any OTel-instrumented service, collector or tool
Coding Agent TelemetryClaude Code, GitHub Copilot CLI, OpenAI Codex CLI
Trace ConnectorsAgents hosted in AWS Bedrock and Google Vertex AI
Budget AIRequests sent through the ANTS LLM gateway are traced automatically

Ingestion is asynchronous. New data usually appears within seconds.

Querying data

In the UI

  • Discover AI > Tracing and Sessions for individual traces and conversations.
  • Gauge AI dashboards for cost, latency and usage: Cost Dashboard, Latency Dashboard, AI Usage, Cost Performance Metrics and Latency & Analytics.
  • Gauge AI > Users for per-user totals.

See Observability and FinOps.

With the public API

The REST API under https://<your-ants-host>/api/public/ uses the same Basic authentication as ingestion. It includes:

EndpointReturns
GET /api/public/traces, GET /api/public/traces/{traceId}Traces
GET /api/public/observations, GET /api/public/observations/{observationId}Observations
GET /api/public/sessions, GET /api/public/sessions/{sessionId}Sessions
GET /api/public/metricsAggregated metrics, from a JSON query parameter with view, metrics, dimensions and filters
GET /api/public/v2/scores, POST /api/public/scoresScores
GET /api/public/comments, POST /api/public/commentsComments

The trace list leaves out input and output by default. See the API Reference.

ℹ

A deployment can limit project keys to ingestion only. On such deployments, read requests with a project key return 403.

Retention

Data is kept for the period set under Organization Settings > Data Retention: 30 days, 90 days, 6 months, 1 year, 2 years, 7 years or Indefinite. A nightly job deletes older data. Changing the period needs the data retention entitlement. See Data Retention.

Access control

Who can list traces, open trace content, see per-user data or change settings depends on role. See Who can see traces and Members & Roles.

Next steps

© 2026 ANTS Platform, Inc.Docs v1.0 · Last updated June 2026