@agentprodready/agent-framework

TypeScript agents you can ship this week — with a clean path to production controls when you need them.

Production-oriented architecture with a young ecosystem.

Simple Agent API: createAgent · createTeam · createWorkflow · createOrchestrator · handoff · reference() · openai() · openaiCompatible() · anthropic() · gemini() · tool() · inMemory() · fileMemory() · postgresMemory() · invoke() · stream() · replayStream() · approve() · reject() · resume() · close()

What this package is: an agent + team + Runtime execution entrance — not a graph DSL. Full evaluator FAQ: What is AgentProdReady?.

| Question | Short answer | |---|---| | Core abstraction | createAgent / createTeam over Runtime — not LangGraph-style graphs | | Durable state | Runtime checkpoints; fileMemory / postgresMemory; HITL park/resume; stream replay (memory: true stays ephemeral) | | Retries / idempotency / HITL | Runtime owns retries; tool idempotency + ledger; approve / reject / resume | | Provider routing | Simple helpers pick one model; hosts use Capability Resolution failover | | “Production ready” | Architecture for production controls; young ecosystem — not a huge-fleet claim |

Guides: What is AgentProdReady? · Getting Started · Simple Diagnostics · Anthropic · Gemini · OpenAI-compatible · Simple Tools · Simple Memory · Durable Memory · HITL Approval · Stream Replay

Star / contribute: ameenmari/agentprodready

Install

npm install @agentprodready/agent-framework

Requires Node.js >=22 <25 and an ESM project.

Scaffold:

npm create agentprodready@latest my-agent

Quick start (zero secrets)

import { createAgent, reference } from "@agentprodready/agent-framework";

const agent = createAgent({
  model: reference(),
  instructions: "You are a helpful assistant.",
});

const result = await agent.invoke("Hello");
console.log(result.text);
await agent.close();

No API key. No database. No Docker.

OpenAI

npm install @agentprodready/agent-framework @agentprodready/ai-provider-openai

import { createAgent, openai } from "@agentprodready/agent-framework";

const agent = createAgent({
  model: openai("gpt-4o-mini"),
  instructions: "You are a helpful assistant.",
});

const result = await agent.invoke("Hello");
console.log(result.text);
await agent.close();

Set OPENAI_API_KEY in the environment (the library does not load .env files).

OpenAI-compatible

import { createAgent, openaiCompatible } from "@agentprodready/agent-framework";

const agent = createAgent({
  model: openaiCompatible({
    baseUrl: "https://api.example.com/v1",
    model: "llama-3.1-70b",
  }),
  instructions: "You are a helpful assistant.",
});

Capability id openai-compatible-ai. Credentials: OPENAI_COMPATIBLE_API_KEY (never silent OPENAI_API_KEY fallback).

Anthropic

npm install @agentprodready/agent-framework @agentprodready/ai-provider-anthropic

import { createAgent, anthropic } from "@agentprodready/agent-framework";

const agent = createAgent({
  model: anthropic("claude-sonnet-4-20250514"),
  instructions: "You are a helpful assistant.",
});

Set ANTHROPIC_API_KEY. Messages API — not openaiCompatible().

Gemini

npm install @agentprodready/agent-framework @agentprodready/ai-provider-gemini

import { createAgent, gemini } from "@agentprodready/agent-framework";

const agent = createAgent({
  model: gemini("gemini-2.0-flash"),
  instructions: "You are a helpful assistant.",
});

Set GEMINI_API_KEY. Native Generative Language API — not openaiCompatible().

Tools

import { createAgent, reference, tool } from "@agentprodready/agent-framework";

const agent = createAgent({
  model: reference(),
  instructions: "You are helpful.",
  tools: [
    tool({
      name: "getWeather",
      description: "Get weather for a city",
      parameters: {
        type: "object",
        properties: { city: { type: "string" } },
        required: ["city"],
      },
      execute: async ({ city }) => ({ city, forecast: "sunny" }),
    }),
  ],
});

const result = await agent.invoke('USE_TOOL:getWeather:{"city":"Paris"}');
console.log(result.text);
await agent.close();

Defaults are conservative (mutating / non-idempotent). See the Simple Tools guide.

Memory

Ephemeral: memory: true ≡ inMemory() — process-local, cleared on exit.
Durable (v1.6): fileMemory({ directory }), postgresMemory({ connectionString }) — survives restart.

The reference provider is deterministic and does not perform natural-language reasoning over recalled memory. Use openai() for NL recall demos. See Simple Memory and Durable Memory.

Streaming & replay

for await (const event of agent.stream("Hello", { resumeFrom: 0 })) {
  if (event.type === "text") process.stdout.write(event.text);
}

for await (const event of agent.replayStream(executionId)) {
  // log-only replay
}

Embedded library stream — not HTTP SSE. See Stream Replay.

HITL

approve(approvalId), reject(approvalId), resume(executionId) for approval-required tools. See HITL Approval.

Diagnostics

Successful invoke results include result.metadata (provider, modelId, durationMs, tools counts, optional memory). See Simple Diagnostics.

Production path

When you outgrow the weekend path: embed deployment · production deployment · adopting.

Simple/embedded mode is not production HTTP authentication.

Advanced APIs

This package also exports the advanced Agent Framework. Simple helpers do not deprecate it. Production multi-tenant hosts should use advanced Security, Runtime, Composition, and durable Memory/Persistence as documented in the repository.