ANTS Platform Documentation

Your guide to building, governing, and scaling AI agents in the enterprise — from first SDK call to your auditor's signature.

What is ANTS Platform?

ANTS is a comprehensive AI Governance Platform (AIGP) that pairs agentic observability with built-in security and regulatory compliance. It's the control plane for the AI workforce running across your enterprise — whether you sanctioned it or not.

A built-in Discovery & Visibility layer inventories every agent, model, tool, and data lineage running in your org. LLMOps serves as the foundation — powering AI Governance with integrated cost intelligence and resilient runtime controls.

Platform Overview

ANTS is organized into three concentric layers. Each layer is independent — you can adopt them in any order — but they're designed to compose into a single audit-ready control plane.

LLMOps Framework

The operational core. Captures every agent, prompt, model, and decision in real time, with the same primitives you'd expect from a mature APM stack — but native to non-deterministic systems. Supports OpenAI, Anthropic, AWS Bedrock, Azure, Google Vertex, Gemini, Cohere, Mistral, Groq, LangChain, and LlamaIndex out of the box.

Observability in LLM Applications

OpenTelemetry-native distributed tracing for chains, agents, retrievals, and tool calls. Replay any run, attach any artifact, diff any prompt — across providers. Two integration paths: auto-instrumentation (~10 lines, zero changes to existing code) or the full SDK for custom spans and metadata.

quickstart.pycopy
from ants_sdk import AntsPlatformSpanProcessor
from opentelemetry.sdk.trace import TracerProvider

class=class="tok-s">"tok-c"># Initialize ANTS in two lines
provider = TracerProvider()
provider.add_span_processor(AntsPlatformSpanProcessor(
    public_key=class="tok-s">"pk-ap-...",
    secret_key=class="tok-s">"sk-ap-..."
))

@trace(project=class="tok-s">"support-agent")
def answer(question: str) -> str:
    class=class="tok-s">"tok-c"># Every LLM call inside is automatically traced
    return llm.complete(prompt=PROMPTS.support_v7, q=question)

Prompt Management

Version, A/B-test, and roll back prompts without redeploying. Compare cost, latency, and quality between candidates in one view. Prompts are first-class objects with full audit trails — every change, every author, every downstream impact tracked.

Evaluation

Run LLM-as-judge, classifiers, and your own rubrics on production traffic. Batch over datasets to gate every release in CI — ANTS returns a structured pass/fail report your pipeline can act on. Credit consumption: 2 credits per 1,000 evaluation runs.

AI Governance

The control layer. Runtime policies enforce what's allowed, evidence packs satisfy your auditors, and Shadow AI discovery surfaces the agents nobody told you about.

• Shadow AI discovery — every unsanctioned model, agent, and tool, scored by risk. Browser extension + desktop agent; supports ChatGPT, Claude, Copilot, Gemini, GitHub Copilot, Cursor.
• Runtime policy engine — PII, PHI, PCI, jailbreak, and unsafe-tool blocking at the edge of every model call. Policies checked in <8ms.
• Evidence & audit trail — exportable artifacts for SOC 2, ISO 42001, EU AI Act, and NIST AI RMF. RBAC with Owner, AI Steward, and AI Developer roles.

Popular Integrations

Auto-instrumentation for every major LLM provider, framework, and ops tool — or write five lines of SDK and you're done. Browse the full integrations catalog.

Support & Community

Engineering questions go to GitHub Discussions; operational questions go to your dedicated Slack channel (Team and Enterprise); architecture and compliance questions go to your named solutions engineer.

What's Next?

1. Install the SDK and trace your first agent — five minutes.
2. Wire your CI to run dataset evaluations on every PR.
3. Turn on Shadow AI discovery to see what's actually running in your org.
4. Promote a guardrail from shadow to enforce mode.
5. Export your first SOC 2 evidence pack — send it to your auditor.