The production layer

Everything an AI agent needs around the SDK you already use

Your orchestration stays yours. agentFast owns the layer above it.

agentFast architecture: the production layer sits between your agent SDK and your productFour layers. L1 orchestration (LangGraph, CrewAI, Claude Agent SDK, OpenAI Agents SDK, Vanilla) stays your own idiomatic code. L2 is a ~200-LOC adapter seam. L3 is the shared SDK-agnostic production layer — memory, tools, observability, guardrails, human-in-the-loop and evals. L4 is the product surface.L4Product surfaceplayground · dashboard · HITL UI · CLIL3Production layermemory · tools · observability · guardrails · HITL · evalsL2Adapter seam~200 LOC per SDK · checkpoint · resumeL1Orchestration — your code, untouched← the productLangGraphCrewAIClaude Agent SDKOpenAI Agents SDKVanilla

L3 — Capabilities

Durable Execution for AI AgentsSurvives SIGKILL. Resumes at the exact step. No replay.Human-in-the-Loop for AI AgentsDurable approval gates. Exactly-once side effects.AI Agent ObservabilityStep tree, token cost, latency. OTEL + LangSmith.AI Agent GuardrailsPII redacted pre-model. Injection, abuse, rate limits.AI Agent MemoryShort-term, long-term, episodic, scratchpad + compaction.AI Agent Evals in CIFresh-context judge. Fails CI on a regression.MCP Tools for AI Agents50+ MCP-native tools. Add any server by config.

L1 — Plugs into your SDK

LangGraph in Production
Native interrupt + checkpointer
~200 LOC adapter →
CrewAI in Production
Real BaseTool bridge
~200 LOC adapter →
Claude Agent SDK in Production
MCP tool bridge
~200 LOC adapter →
OpenAI Agents SDK in Production
Real FunctionTool + Runner
~200 LOC adapter →

Working reference agents

Production AI Support Agent
KB → answer → escalate → resolve
Production AI Research Agent
search → synthesize → cite → export
Production AI Coding Agent
read repo → fix → PR → test
Own the production layer on day one.

Source-available. Download once, run it anywhere, own it for life.

Get agentFast — from $89