Build a team of AI agents. Ship in 6 hours.
Forge Sprint 01 taught you to wield Claude Code as a solo agent. Sprint 02 levels up: you will build a team of AI agents that coordinate, communicate and ship code together, the same way production engineering teams have scaled from 15 pull requests per month to 1,000 using this exact methodology. You bring the architecture, the agent configuration and the skills you built before Sprint day.
· Editions
Forge 2 runs as editions of the same challenge. Pick the edition you want. Open editions link straight to registration, completed editions link to their results.
· Leaderboard
Scores only, the way Forge runs. Switch between the all-time best across every Forge 2 edition and a single edition ranking.
01 The Brief
Build a multi-agent AI system. Ship in 6 hours.
You will use Hermes Agent as your orchestrator, the "Product Owner" agent that plans, decomposes tasks and assigns work, and OpenClaw as your coding agent, the hands that write, test and deploy code. You will wire them together through Slack so every agent-to-agent conversation is visible, auditable and human-in-the-loop.
This is not a toy demo. This is the exact methodology used in production by companies running 2,000+ automated campaigns with 4 named AI agents coordinating through 30+ Slack channels.
The exact challenge brief and dataset are revealed on Sprint morning. You bring the architecture, the agent configuration and the skills you built during your pre-sprint study.
02 What You Will Build
Your submission must demonstrate a working multi-agent system with these five components.
Orchestrator Agent (Hermes)
Acts as Product Owner. Decomposes the challenge into tasks, assigns them to worker agents, tracks progress and reports structured status.
Coding Agent (OpenClaw)
Receives tasks from the orchestrator, writes code, runs tests and pushes to the shared repository.
Slack Communication Layer
All agent-to-agent communication happens through Slack channels. Agents never communicate directly. You, the human, monitor and can intervene at any point.
CI/CD Pipeline
Automated tests run on every agent-generated commit. Only passing code reaches the release candidate stage. Final merge to main branch is human-approved.
Canary or Health Check
Some mechanism to verify the output quality, such as test pass rate, accuracy score or error classification.
- · Model routing: using an expensive model (Opus/GPT-4) for planning and a cheap model (DeepSeek/GLM/Nemo) for execution.
- · Self-improving skills: agents that learn from errors and create reusable skills.
- · Structured status reporting: the "What I Did / What's Left / What Needs Your Call" pattern.
- · Error classification and auto-fix loops.
03 Example Problem Domains
The actual challenge is revealed on Sprint morning. These examples show the flavour.
Multi-Agent SEO Audit Pipeline
Orchestrator decomposes a website audit into 5 parallel tasks: technical SEO, content quality, schema validation, performance and backlinks. Each task is assigned to a specialist sub-agent. Results are merged into a unified report.
Automated Content Factory
Orchestrator reads a content brief, assigns research to Agent A, outline to Agent B, first draft to Agent C and formatting/export to Agent D. Human reviews the final output.
Competitive Intelligence System
One agent scrapes competitor websites, another analyzes pricing, a third monitors social mentions and the orchestrator synthesizes a weekly competitive brief.
Bug Triage + Auto-Fix
Feed the system a list of GitHub issues. Orchestrator prioritizes by severity, assigns to coding agents, agents write fixes, CI runs tests and the human reviews PRs.
Your own idea
Build any multi-agent system that solves a real business problem. Surprise us.
04 The Stack
Required Tools
| Tool | Role | Why |
|---|---|---|
 Hermes Agent | Orchestrator | Open-source, self-improving, persistent memory, skill learning. The "brain" that coordinates everything. |
 OpenClaw | Coding Agent | Multi-channel gateway connecting chat apps to LLM coding agents. The "hands" that write code. |
| Slack (free workspace) | Communication | All agent communication happens here. Human-in-the-loop. Matches production patterns. |
| GitHub | Version Control | Public repo with incremental commits. Required for judging. |
Model Options (All Free)
| Model | Provider | Best For | Free Tier |
|---|---|---|---|
| Kimi K2 Instruct | Groq | General coding + orchestration | ~1,000 req/day |
| DeepSeek V3 | DeepSeek API | Bulk coding tasks | Free tier available |
| Qwen2.5-Coder 7B | Ollama (local) | Offline coding | Unlimited (local) |
| GLM-4 | Zhipu AI | Task coordination / PO role | Free tier available |
Model routing is encouraged. Use a stronger model for planning (Orchestrator) and a cheaper model for execution (Coding Agent). This mirrors the production pattern where Opus architects and DeepSeek executes.
05 Forge Roadmap Connection
| Sprint | Theme | Date | Connection |
|---|---|---|---|
| Forge 1 | Claude Code (Solo Agent) | Jun 6 | Foundation: learn to wield a single AI coding agent |
| Forge 2 | Hermes + OpenClaw (Multi-Agent) | Jun 20 | Level up: orchestrate a TEAM of AI agents |
| Forge 3 | ProcessHQ Hackathon | Jul 4 | Apply: use your agent team to build real product features |
Forge 2 to 3 pipeline: the top performers from Forge 2 are automatically invited to Forge 3 (ProcessHQ Hackathon), where they will use the multi-agent skills they built here to solve real product development challenges for ProcessHQ.AI, a live SaaS product with 56+ enterprise demos and paying customers.
06 Sprint Day Schedule, 20 June 2026
| Time (IST) | Block | Details |
|---|---|---|
| 09:00 to 09:30 | Check-in | Photo ID, laptop setup, Slack workspace join, stack verification |
| 09:30 to 09:50 | Ignition Briefing | Welcome, Multi-Agent 10 Rules methodology overview, judging rubric |
| 09:50 | Challenge data released | Real dataset + problem brief published |
| 10:00 | Build window opens | 6-hour countdown. Mentors available throughout. |
| 10:30 | First checkpoint | Verify Hermes + OpenClaw + Slack are connected and communicating |
| 13:00 to 14:00 | Catered lunch | Builders, mentors and judges eat together |
| 14:00 | Architecture check-in | Forge Captain reviews agent configurations and Slack channel structures |
| 15:30 | Scope-or-cut | Stabilize core pipeline. Drop anything that will not ship. |
| 16:30 | 30-minute warning | Final testing. Record Loom walkthrough. |
| 17:00 | Submission lock | GitHub commit frozen. Surprise second dataset released to judges. |
| 17:00 to 17:30 | Auto-tester run | Judges run every build against the surprise dataset |
| 17:30 to 19:00 | Finalist demos | 5 minutes each: 3-min demo + 2-min Q&A from judges |
| 19:00 | Judging deliberation | |
| 19:30 | Champions announced | Prizes + Forge 3 invitations extended on the night |
07 Judging
Scoring rubric, 100 points total.
| Criterion | Points | What Judges Look For |
|---|---|---|
| Agent Architecture | 25 | Clean separation of orchestrator and worker agents. Hermes manages, OpenClaw executes. Clear role boundaries. |
| Slack Communication | 20 | All agent communication visible in Slack. Structured status updates. Human can intervene at any point. No direct agent-to-agent calls outside Slack. |
| Working Output | 20 | The system actually produces a correct, useful result against the challenge dataset. Quality of the final deliverable. |
| Model Routing | 10 | Strategic use of different models for different tasks. Expensive for planning, cheap for execution. Demonstrated cost awareness. |
| CI/CD & Quality Gates | 10 | Automated tests on agent output. Some form of quality verification before final output. Human approval gate for final merge. |
| Self-Improvement | 10 | Agents that learn from errors, create reusable skills or improve across iterations. Hermes learning loop in action. |
| Code Quality & Docs | 5 | Clean repo, meaningful commits, README, agent-log.md preserved. Can explain any part of the code. |
- Free-stack builders (Groq/Ollama) win ties over paid-subscription builders.
- If still tied: the system that handles the surprise dataset better wins.
- If still tied: code reading score decides.
08 Prizes & Outcomes
What Every Participant Gets
| Outcome | Details |
|---|---|
| Cash Prizes | Revealing soon, to be announced before Sprint Day |
| NMG Offer | Top performers receive NMG apprenticeship, internship or full-time offers |
| Certificate of Participation | Official nmg.labs Forge Sprint 02 certificate for all participants |
| NMG Labs Skool Membership | Lifetime membership to the NMG Labs Skool community: private builder network, resources, mentorship and priority access to future Forges |
Place-Specific Outcomes
| Place | Prize | Forge 3 Access |
|---|---|---|
| Champion (1st) | Cash prize (revealing soon) + NMG offer + Certificate + NMG Labs Skool | Guaranteed seat + dedicated NMG mentor for Forge 3 |
| Runner-up (2nd) | Cash prize (revealing soon) + NMG offer + Certificate + NMG Labs Skool | Guaranteed seat at Forge 3 |
| 3rd Place | Cash prize (revealing soon) + NMG offer + Certificate + NMG Labs Skool | Guaranteed seat at Forge 3 |
| Top 10 Finalists | Certificate + NMG Labs Skool membership | Priority registration for Forge 3 |
| All Participants | Certificate of Participation + NMG Labs Skool membership | Can apply for Forge 3 |
The real prize: Forge 3 (ProcessHQ Hackathon)
Forge 2 is the qualifying round. The skills you build here, multi-agent orchestration, Slack coordination, model routing and CI/CD pipelines, are exactly what you need for Forge 3, where you will build real features for ProcessHQ.AI, a live SaaS product with 56+ enterprise demos and paying customers.
09 Rules
- · Open to 3rd/4th year UG, PG and recent graduates (within 12 months) across India.
- · Forge 1 attendance is NOT required. Fresh builders welcome.
- · Individual competition. Solo sprint.
- Hermes Agent must be the orchestration layer.
- OpenClaw must be at least one of the coding agents.
- Slack must be the communication channel (free workspace is fine).
- Additional agents, tools and frameworks are permitted on top.
- Public GitHub repo with incremental commits required.
- Public GitHub repo with agent-log.md (unedited Hermes/OpenClaw logs)
- Slack export or screenshots showing agent communication
- README documenting architecture (which agent does what)
- Architecture diagram (Mermaid, hand-drawn or any format)
- Loom video walkthrough (max 5 minutes)
- Working demo that can run against the surprise dataset
- Private repo at deadline
- Missing agent-log.md or Slack evidence
- No Hermes Agent in the architecture (using only Claude Code = wrong sprint)
- Single giant commit (no build phases)
- Auto-tester failure within 5 minutes
- Cannot explain any block of your own code
- 75%+ code duplication with existing public repos
- Conduct violation
Intellectual property: your code, your IP. NMG retains a non-exclusive license to publish a case study about your build with credit to you, and you are free to commercialize, open-source or fork your project after.
Code of conduct: zero tolerance for plagiarism, harassment or AI-generated submissions misrepresented as your own work. Full details in the Terms & Conditions and Code of Conduct.
Lock your seat for Forge 2 · Edition 1.
Forge 2 is an individual competition. Configure your agents, bring your skills and register below. We will email you the check-in details before Sprint day.
Forge Captain · Ayush Gupta · labs@nmgdigital.com