Forge-AI Protocol v3.1

18 AI Agents. 8 Phases.
One Startup a Day.

Forge-AI is an autonomous multi-agent pipeline that runs every day at 00:00 UTC. It detects emerging tech trends, designs a micro-startup concept, writes a production-ready Spring Boot application, validates it compiles, enriches it with 17 specialist AI agents, and publishes the result here.

No human writes the code. No human writes the pitch. The pipeline handles everything — from market sizing to EU AI Act compliance — for roughly $0.005 per run.

18

AI Agents

8

Pipeline Phases

~$0.005

Cost per Run

20+

Output Files

The Pipeline

[01]

Trend Detection

parallel

Parallel Scout × 3

Three ScoutAgents scan the web simultaneously via Tavily, detecting the strongest emerging tech trend of the day. The best signal wins.

ParallelScoutService
Tavily Web Search

Outputs

  • trends.json
  • 3 candidate topics
[02]

Business Case

ProductManagerAgent

The PM distills the winning trend into a full BusinessCase: startup name, problem statement, target audience, monetization angle, MVP features, and Spring Boot starters.

ProductManagerAgent

Outputs

  • BusinessCase.json
[03]

Parallel Incubation

parallel

4 agents in StructuredTaskScope

The heaviest phase. Claude Opus generates a complete Spring Boot 4 project with all MVP features while three other agents run in parallel: market sizing, growth strategy, and cloud cost estimation. Java 25 Virtual Threads + StructuredTaskScope with a 45-minute timeout.

ForgeEngineer
MarketAnalystAgent
GrowthAgent
FinOpsAgent

Outputs

  • Full Spring Boot source code
  • MarketAnalysis.json
  • MarketingMaterial.json
  • CostEstimate.json
[04]

Maven Validation + Auto-Healing

Up to 3 attempts

The generated code is compiled. On failure, ValidatorAgent (Claude) performs a surgical analysis of Maven errors and generates targeted fix prompts. ForgeEngineer corrects only the broken files — no full regen.

MavenValidatorTool
ValidatorAgent

Outputs

  • BUILD SUCCESS or auto-healed project
[05]

Enrichment

parallel

4 agents in parallel

Claude writes JUnit 5 test suites directly to disk. Ethics Officer checks EU AI Act compliance. TechWriter generates an OpenAPI 3.1 spec with curl examples. SharkTank delivers a brutal code + market review.

EthicOfficerAgent
QAEngineerAgent
TechnicalWriterAgent
SharkTankAgent

Outputs

  • ETHICS_GUIDE.md
  • src/test/java/*.java
  • openapi-spec.yml
  • CURL_EXAMPLES.md
  • SHARK_TANK_REVIEW.md
[05.5]

Simulation

parallel

5 agents in parallel

Five agents simulate real-world scenarios: UX testing with 3 personas, SRE deployment planning with real cloud pricing, IP patent analysis, VC investment scoring, and cross-project synergy detection across the entire solutions/ catalog.

UserSimulatorAgent
SREAgent
IPGuardAgent
VCAgent
EcosystemWeaverAgent

Outputs

  • deploy/Dockerfile
  • deploy/k8s-hpa.yml
  • SLO_REPORT.md
  • IP_STRATEGY.md
  • VC_ASSESSMENT.md
  • ECOSYSTEM_SYNERGY.md
[06]

CFO Report

CFOAgent

The CFO receives all financial signals — FinOps cost estimate, real LLM token costs from this run, market analysis — and produces unit economics: CAC, LTV, payback period, Y1-Y3 ARR projections.

CFOAgent

Outputs

  • FINANCIAL_REPORT.md
[07–08]

Pitch + Chronicle + Publish

PitchAgent + ChronicleAgent

PitchAgent synthesizes all 17 agent outputs into a 30-second elevator pitch. ChronicleAgent writes a 1,500-word narrative article for The Forge Chronicles. The startup report is persisted, the README is updated, and Telegram notifications are sent.

PitchAgent
ChronicleAgent

Outputs

  • startup-report.json
  • MEETING_LOG.json
  • CHRONICLE.md
  • Telegram notification
  • README update

Model Architecture

Each agent is assigned to one of three model tiers based on the complexity of its task. This keeps costs minimal without sacrificing quality where it matters.

CRITICAL

Claude Opus 4.6

Code generation, test writing, error healing. The architect.

Only Claude can generate 1,000+ line Spring Boot projects that compile on the first try.

ForgeEngineer
QAEngineerAgent
ValidatorAgent
STRATEGIC

Gemini 2.5 Pro

Market analysis, investment scoring, legal. The analysts.

Deep reasoning over large context windows. Used where quality of analysis matters more than speed.

ProductManagerAgent
MarketAnalystAgent
SharkTankAgent
VCAgent
IPGuardAgent
EthicOfficerAgent
FAST

Gemini 2.5 Flash

Scouts, narrators, simulators, pitch. High-volume tasks.

10× cheaper, still excellent. Used for tasks where breadth matters more than depth.

ScoutAgent
GrowthAgent
FinOpsAgent
PitchAgent
UserSimulatorAgent
SREAgent
TechnicalWriterAgent
EcosystemWeaverAgent
CFOAgent
ChronicleAgent

Tech Stack

Runtime

Java 25Virtual Threads + StructuredTaskScope

Parallel agents with structured concurrency

LangChain4j 1.12.2AiServices + RetryingChatModel

Agent framework with auto-JSON deserialization

Jackson 2.xRecord-based POJOs

All agent outputs are strongly-typed Java records

AI Models

Claude Opus 4.6via Claude Code CLI

Code generation — runs as a subprocess

Gemini 2.5 Provia Google AI SDK

Strategic analysis with exponential backoff

Gemini 2.5 Flashvia Google AI SDK

High-throughput tasks, wrapped in RetryingChatModel

TavilyWebSearchTool

Live web search for trend detection

Generated Startups

Spring Boot 4.xGenerated from scratch per run

Scaffolded by SpringInitializrTool

Spring AILangChain4j or Spring AI

AI library chosen by ProductManagerAgent

MavenValidated with mvn clean compile

Auto-healed on failure by ValidatorAgent

Frontend & Infra

Next.js 15App Router + Static Generation

This site — reads solutions/ at build time

VercelEdge deployment

Auto-deploys on push to main

Tailwind CSS 4Dark neon aesthetic

CRT scanlines, star field, neon glow

Java 25 Concurrency Model

The pipeline uses Java's structured concurrency primitives to run agents in parallel with strict lifetime guarantees. All forks are scoped — if one fails, the scope handles it cleanly.

// Phase 3: 4 agents running in parallel

try (var scope = StructuredTaskScope.open(

Joiner.awaitAll(), cfg => cfg.withTimeout(Duration.ofMinutes(45)))) {

scope.fork(() => agents.engineer.forge(...));

scope.fork(() => agents.marketAnalyst.analyze(...));

scope.fork(() => agents.growthAgent.strategize(...));

scope.fork(() => agents.finOps.estimate(...));

scope.join();

}

Output per Startup

Every forged startup is a self-contained directory with 20+ files. Here's what gets generated automatically in a single pipeline run.

src/main/java/**/*.java

Full Spring Boot application with all MVP features

pom.xml

Maven build — Spring Boot 4 + Spring AI + Jakarta EE

startup-report.json

All 17 agent scores aggregated into one JSON

EXPLAINER.md

Technical architecture doc written by ForgeEngineer

SHARK_TANK_REVIEW.md

Cross-model code + market review (Gemini → Claude)

VC_ASSESSMENT.md

Investment thesis, funding stage recommendation

FINANCIAL_REPORT.md

Unit economics, CAC/LTV, Y1-Y3 ARR projections

ETHICS_GUIDE.md

EU AI Act compliance, risk category, ethical guidelines

MARKETING_PLAN.md

Growth strategy, tagline, domain suggestions

openapi-spec.yml

OpenAPI 3.1 spec for all generated endpoints

CURL_EXAMPLES.md

Ready-to-run curl commands for every endpoint

IP_STRATEGY.md

Patentable claims, license analysis

SLO_REPORT.md

Service Level Objectives for production

deploy/Dockerfile

Optimised multi-stage Docker image

deploy/k8s-hpa.yml

Kubernetes HPA + deployment manifest

ECOSYSTEM_SYNERGY.md

Cross-project integration opportunities

CHRONICLE.md

1,500-word narrative article for The Forge Chronicles

MEETING_LOG.json

Full board meeting transcript — all 17 agents

AGENT_LOG.md

Incremental per-phase progress log

Why

Forge-AI is an experiment in autonomous software generation. The question is simple: how far can a pipeline of coordinated AI agents go without human intervention?

Each day, the pipeline starts from nothing — no pre-written code, no templates, no manual input. It finds a trend, designs a product, writes the software, validates it compiles, reviews it from six different professional perspectives, and prices it for acquisition — all autonomously.

The cross-model review is deliberate. Claude generates the code; Gemini reviews it as a SharkTank investor. Neither model is reviewing its own output. This adversarial setup catches what self-review misses.

The startups listed on this site are real generated artifacts — functional Spring Boot projects, not demos. The scores are real agent outputs. The financial reports are real model reasoning about real cost data.

See it in action

Browse today's forged startup

→ Open Alpha Stream