Strata AI SDK

Core contract layer for AI agents. Runtime-agnostic. Enterprise-ready by default.

📦 Installation

uv add strata-ai

🧩 Module Structure

🔁 Runtime Inversion

Agent patterns never import framework internals. The runtime is injected:

from strata_ai import ReActAgent, AgentConfig
from strata_ai.runtime import LangGraphAdapter

agent = ReActAgent(
    config=AgentConfig(name="analyst", model="openai:gpt-4o", instructions="..."),
    runtime=LangGraphAdapter(checkpointer=postgres),
    tools=[sql_query, web_search]
)
result = await agent.run({"question": "What is the GDP of Kenya?"})

Swap runtimes in config. Zero pattern refactors.

🛡️ Enterprise Defaults

• Observability: OTel spans auto-emitted on every run(), stream(), tool.call. Ships to MLflow/Langfuse/OTLP.
• Security: PIIFilter (MSISDN, ID, email, refs) + Guardrail middleware. Enabled when pii_sensitive=True.
• Governance: AuditLog (immutable) + LineageTracker (MLflow-backed). GDPR-compliant sink adapters.
• Errors: StrataBaseError hierarchy → RFC 9457 ProblemDetail at API edge. Zero 500 leaks.

🧪 Testability

from strata_ai.runtime import MockRuntime

# Strict mode enforces 5 invariants: state immutability, tool-result pairing, checkpoint-before-resume, audit-on-violation, deterministic threads
agent = ReActAgent(config=..., runtime=MockRuntime(strict=True), tools=[...])
result = await agent.run({"input": "test"})

Zero LLM calls. Fast CI. Pattern parity verified via tests/integration/test_runtime_parity.py.