Turn a Single Sentence into a Production‑Ready AI Agent with AgentForge

Problem

Four user groups face distinct obstacles when building AI agents:

Non‑technical users : need AI automation (e.g., customer routing, on‑call broadcasting, itinerary planning) but cannot write code, design prompts, or handle container deployment.

Front‑end/Node engineers : must recreate Dockerfiles, Aone container adapters, Egg startup scripts, scheduler integration, DingTalk hooks, and memory stores for every new agent, leading to duplicated effort.

Java developers : core Java teams lack a Node/Python‑based agent ecosystem; building a Java agent requires learning a new stack or maintaining a fragile custom framework.

Small teams : need a way to share agents without rewriting code for each use case.

Solution – AgentForge Platform

AgentForge is an “Agent factory” that lets a user describe a capability in a single sentence. The Builder streams a 7‑layer prompt (SOUL, USER, AGENTS, SKILL_CHAIN, TOOLS, MEMORY, etc.), automatically matches required skills and MCP equipment, and produces a production‑ready agent.

Workflow for Non‑Technical Users (5‑minute Loop Agent)

Describe the desired capability in one sentence.

Generate the full prompt stack and skill matching via the Builder.

Upload a ZIP or write a Markdown skill; the platform validates and sandbox‑checks it.

Debug using a dual‑mode UI that toggles between equipment view and dialogue view, showing SSE events for routing, tool calls, and memory retrieval.

Deploy with one click; the cloud runs a true ReAct‑style loop with function calling, memory recall, and tool confirmation/retry.

Support for Node Engineers

Complete Dockerfile and Aone container adaptation (APP‑META/docker‑config, nodejsctl start_app, health_check, setenv).

Full SOP covering BUC authentication → MySQL/Redis config → Egg startup → migration → health check.

ANVIL runtime for single agents and ANVIL‑MULTI for multi‑agent scenarios, both running inside Aone containers.

Support for Java Developers

AgentForge appears as an HTTP service; no Node code is required.

Build an agent in the browser (same 5‑minute flow) and obtain a deployable instance.

Three integration channels:

HTTP API – POST /api/chat with SSE streaming.

DingTalk robot – zero‑code interaction via group mentions.

MCP reverse call – expose a Java service as an MCP server; business methods become tools.

Cross‑session shared memory automatically accumulates context.

Zero intrusion: no additional Maven dependencies, no LLM SDK keys, no changes to existing pom files.

Team Collaboration – Agent Teams

All agents share a unified contract (persona, capabilities, message handling, result reporting).

TeamCanvas visual canvas enables drag‑and‑drop composition of Manager‑Worker teams without code.

Protocol‑driven role communication allows composable teams (e.g., customer‑service + itinerary planning, or dev‑task dispatch + code‑review + doc generation).

Shared memory namespaces prevent duplicate knowledge while preserving private skills.

Core Technical Assets

fliggy‑memory‑sdk – a self‑built mem0‑compatible long‑term memory SDK with namespace isolation, catalog injection, top‑K retrieval, fact extraction, decay/merge/deduplication.

FECHO – internal HSF‑to‑MCP gateway that automatically converts any HSF service into an MCP endpoint, giving agents access to the entire corporate service catalog.

ANVIL / ANVIL‑MULTI – custom agent runtime frameworks (single and multi‑agent) providing a full‑link production loop.

chatLoop – 2,200‑line ReAct + Function Calling loop exposing SSE events, tool confirmation/retry, history compression, and error handling.

7‑Layer Prompt Templates – separate editing, versioning, and reuse of SOUL, USER, AGENTS, SKILL, SKILL_CHAIN, TOOLS, MEMORY prompts.

Engineering Practices

Five‑stage agent lifecycle (Generate → Equip → Debug → Deploy → Schedule) with independent contracts for each stage.

Multi‑environment consistency via config.{env}.ts and Aone‑derived EGG_SERVER_ENV; automatic migration; zero extra config in production.

DB‑first + file fallback storage for persistent RDS or local MySQL.

Protocol‑defined extension points for Skill, MCP, and Memory, allowing any team to add capabilities without changing platform code.

Results

Platform runs stably and is privately deployed across multiple business lines.

Real agents built include customer‑routing assistants, dev‑task dispatchers, debate judges, on‑call broadcasters, itinerary planners, etc.

End‑to‑end integration chain: BUC authentication → Aone GitLab deployment → SchedulerX scheduling → DingTalk robot → corporate MCP services.

Roadmap

Short term : enable any business user to create a first agent within 5 minutes.

Mid term : allow agents to call each other, forming a Lego‑like agent ecosystem with per‑team ANVIL repositories.

Long term : make “building agents” a routine task for ordinary users and “deploying agents to production” a zero‑setup operation.