A few weeks ago I picked 4 features from my backlog, typed one command, and walked away from my desk.
I made coffee. I did laundry. I checked on the build about 40 minutes later.
Three features were done - branches created, tests written, code implemented, PRs ready for review. The fourth was still in progress, working through a tricky edge case in the payment integration.
No merge conflicts. No "which file did agent 2 break?" debugging. Each agent worked in its own isolated worktree, on its own branch, completely unaware of the others.
This is what I spent the last year building. And it almost didn't happen.
Three months before the coffee moment
Rewind to early 2025. I'm building a B2B platform - NestJS monorepo, TypeScript, the usual stack. I use Claude Code for everything. And every single session starts the same way.
"Here's how my guards work. Here's the interceptor chain. Here's why the repository pattern matters. No, don't put business logic in the controller."
Twenty minutes. Every. Session.
I wrote a CLAUDE.md file. Helped for a week. Then my codebase evolved and the document didn't. The AI would read stale rules and generate code that didn't match the current architecture.
So I built a tool that audits your repo automatically - framework, layers, naming conventions, database patterns, auth setup - and generates a .claude/ directory with rules derived from your actual code. Not a template. Your code, your patterns, your conventions.
The first time I ran it, the AI went from "explain your guards again" to just... following the patterns. No more 20-minute warmups.
That was the first piece of what became Forge DevKit.
Then I caught my AI lying about tests
With architecture rules in place, I asked Claude to add a payment webhook handler. Code looked great. Tests passed. 100% coverage.
Then I read the tests.
Every assertion was testing mock data. The coverage number was real. The verification was fake. The AI had optimized for "green checkmark" instead of "proves the webhook actually updates order status."
I added a rule: tests must trace back to acceptance criteria (AC) - the actual requirements, not the code. If an AC says "webhook updates order status," there must be a test that verifies that exact thing.
The AI started arguing. Not literally, but it generated explanations: "The type system already covers this." "This is an implementation detail." Each one sounded perfectly reasonable.
It was rationalizing. Producing smart-sounding arguments for why it should skip work.
I cataloged these patterns. Found 50+ distinct rationalizations over three months. Built detection for them. When the AI tries to skip a test mapped to a requirement, it gets blocked. It can't argue its way out.
The first time it caught a rationalization I was about to approve myself - that's when I knew this was more than a side project.
From rules to a pipeline to Agent Teams
Rules weren't enough. I needed a full process with gates that block bad output at every step.
So I built a 5-phase TDD pipeline. When you run /forge:dev "Add payment webhook":
- Gate - creates a branch, loads acceptance criteria, links your task tracker
- Test (RED) - generates tests from requirements, not from code
- Implement (GREEN) - writes code to make those tests pass
- Verify - type check, lint, quality patterns, requirement coverage check
- Close - creates PR, updates tracker
The pipeline runs off specs I already wrote - acceptance criteria, architecture rules, quality gates. It doesn't ask me "should I write this test?" It reads the requirements, generates tests, writes code, and only stops when information is genuinely missing. Not hand-holding - actual autonomy with guardrails.
Once I had a pipeline that reliably produced shippable code from a single command... the next question was obvious.
What if I ran multiple pipelines at the same time?
That's the autopilot module. You point it at your backlog, it analyzes which features can run in parallel without file conflicts, groups them into waves, and spawns an AI agent for each feature - each in its own isolated git worktree with its own branch and its own ports.
The agents don't know about each other. They can't step on each other's code. When they're done, each one has a clean PR ready for your review.
That's how I ended up drinking coffee while 4 agents worked my backlog.
Here's what that actually looks like - a real autopilot session, 9 batches, 39 features:
/forge:autopilot
Autopilot - Batch Autonomous Development
Reading backlog and product artifacts...
BACKLOG.md: 39 tasks across P0-P3
PRODUCT-BACKLOG.md: 5 epics, 43 acceptance criteria
Architecture: Python/FastAPI backend + React/TanStack frontend
Starting Phase 0: Triage.
Step 0.1: Feature Collection
| Priority | Features | Tags |
|---|---|---|
| P0 (critical) | 4 | backend, frontend |
| P1 (high) | 15 | backend, frontend, fullstack |
| P2 (medium) | 12 | backend, frontend |
| P3 (low) | 8 | backend |
39 features collected. Launching classification + conflict analysis in parallel.
Step 0.2 + 0.3: Running in parallel
Spawned 2 analysis agents:
- backlog-analyst - classifying all 39 features by complexity, risk, design needs
- conflict-detector - analyzing file overlaps across all feature pairs
741 file-pair combinations being checked.
backlog-analyst complete:
| Complexity | Count | Examples |
|---|---|---|
| Low | 11 | seed data, CRUD endpoints, simple UI |
| Medium | 18 | chat markdown, theme presets, tracker sync |
| High | 10 | workflow engine, auth system, virtual scroll |
Design artifacts needed: 3 features (theme engine, workflow advanced, auth).
conflict-detector complete:
Conflict Matrix Summary
- 682 SAFE pairs (no shared files)
- 53 RISKY pairs (shared files, manageable)
- 6 BLOCKING pairs (same core module, must be sequential)
Key blocking chains:
1. Core models -> Seed -> API endpoints (foundation)
2. Design tokens -> Theme engine (frontend)
3. Workflow core -> Workflow templates -> Marketplace (backend)
These blocking chains define the batch ordering.
Step 0.5: Batch Planning
Based on conflict graph + dependency chains:
| Batch | Theme | Features | Strategy | Dependencies |
|---|---|---|---|---|
| 001 | Core domain models | 4 | 2 parallel | none |
| 002 | Seed data pipeline | 3 | 1 sequential | batch 001 |
| 003 | API endpoints v1 | 5 | 3 parallel | batch 002 |
| 004 | Domain nodes + Tracker | 3 | 2 parallel | batch 002 |
| 005 | Frontend foundations (P0) | 4 | 2 parallel | batch 003 |
| 006 | Frontend polish + design | 7 | 2 waves (3+1) | batch 005 |
| 007 | Workflow engine | 3 | 1 sequential | batch 003 |
| 008 | Agent tools | 5 | 2 parallel | batch 003 |
| 009 | Workflow advanced | 5 | 2 waves (2+1) | batch 007 |
9 batches, 39 features, 20 teammate instances planned.
Batches 003/004 can run in parallel. Batches 007/008 can run in parallel.
Step 0.7: Autonomy Gate
| Check | Status |
|---|---|
| All features have AC | 36/39 (3 inferred from code) |
| Conflict resolution plan | all BLOCKING pairs in same teammate |
| Worktree isolation | each teammate gets own branch |
| Rollback strategy | each batch independently revertable |
Autonomy: full-auto approved.
Generating teammate briefs for all 9 batches... done.
Starting execution. You can walk away now.
going for coffee
Batch 001: Core Domain Models
Spawning 2 teammates in isolated worktrees.
| Teammate | Features | Model |
|---|---|---|
| A | User + Organization models | sonnet |
| B | Crew + Node models | sonnet |
<teammate-message teammate_id="teammate-a" color="blue" summary="User + Org models: 2 models, 2 migrations, membership system">
Done. User + Organization with membership roles. 2 Alembic migrations.
</teammate-message>
<teammate-message teammate_id="teammate-b" color="green" summary="Crew + Node models: categories, knowledge packs, 2 migrations">
Done. Crew with roles, Node with categories + knowledge packs. 2 migrations.
</teammate-message>
Batch 001 complete. 4 features, 15 min, 0 conflicts. 42 tests passing. Starting 002.
Batch 002: Seed Data Pipeline
1 teammate, sequential chain: categories -> knowledge packs -> node definitions.
<teammate-message teammate_id="teammate-a" color="blue" summary="Seed pipeline: 8 categories, 12 knowledge packs, 28 nodes">
Done. Full seed: 8 categories, 12 knowledge packs, 28 node definitions.
</teammate-message>
Batch 002 complete. 3 features, 18 min. GET /api/nodes/ returns 28 nodes. 67 tests. Starting 003 + 004 in parallel.
Batch 003 + 004: Running in Parallel
Batch 003 - API Endpoints: 3 teammates (Crews+Nodes CRUD, Users+Orgs CRUD, Orchestrator)
Batch 004 - Domain Nodes: 2 teammates (Product+QA nodes, Tracker integration)
5 teammates working simultaneously across 2 batches.
<teammate-message teammate_id="batch003-c" color="yellow" summary="Orchestrator: run/stop/status, WebSocket streaming">
Batch 003 Teammate C done. Orchestrator endpoints + WebSocket real-time streaming.
</teammate-message>
<teammate-message teammate_id="batch004-b" color="green" summary="Tracker: config model, GitHub/Linear/ClickUp sync">
Batch 004 Teammate B done. TrackerConfig + external sync (GitHub, Linear, ClickUp).
</teammate-message>
<teammate-message teammate_id="batch003-a" color="blue" summary="Crew + Node CRUD with filtering, pagination, bulk ops">
Batch 003 Teammate A done. Full CRUD with filtering, pagination, bulk operations.
</teammate-message>
<teammate-message teammate_id="batch003-b" color="green" summary="User + Org CRUD, membership, invites">
Batch 003 Teammate B done. User + Org management with roles and invite system.
</teammate-message>
<teammate-message teammate_id="batch004-a" color="blue" summary="Product + QA nodes: 9 definitions, 5 workflows">
Batch 004 Teammate A done. 5 Product nodes + 4 QA nodes with workflows.
</teammate-message>
Batch 003 + 004 complete. 8 features, 23 min, 5 teammates parallel, 0 conflicts. 128 tests. Starting 005.
Batch 005: Frontend Foundations (P0)
2 teammates: app shell+routing+auth (opus) | chat interface+messages (sonnet).
<teammate-message teammate_id="teammate-a" color="blue" summary="App shell: routing, auth flow, sidebar, responsive, dark mode base">
Done. TanStack Router, auth flow, sidebar nav, responsive layout, dark mode.
</teammate-message>
<teammate-message teammate_id="teammate-b" color="green" summary="Chat: message list, rich input, crew panel, WebSocket updates">
Done. Chat interface with message list, command input, crew panel, real-time WebSocket.
</teammate-message>
Batch 005 complete. 4 features, 30 min. Frontend live. 168 tests. Starting 006 (2-wave).
Batch 006: Frontend Polish + Design System (2 waves)
Wave 1: 3 parallel
| Teammate | Features |
|----------|----------|
| A (opus) | Design System: type scale -> colors -> variants |
| B | Chat markdown + virtual scroll |
| C | Animated crew statuses |
Wave 2 (after tokens land): Theme engine + 6 presets
<teammate-message teammate_id="teammate-c" color="yellow" summary="Animated statuses: 4 states, reduced-motion, 220 tests">
Done. 4 animation states, prefers-reduced-motion support. 220 tests.
</teammate-message>
<teammate-message teammate_id="teammate-a" color="blue" summary="Design system: rem types, semantic colors, variants - 74 files migrated">
Done. Type scale (32 files), color tokens (24 files), component variants (18 components). Full foundation.
</teammate-message>
<teammate-message teammate_id="teammate-b" color="green" summary="Markdown rendering + virtual scroll for 10K messages">
Done. Syntax highlighting + virtual scroll handling 10K messages.
</teammate-message>
Wave 1 complete. Design tokens in main. Launching Wave 2: Theme Engine.
<teammate-message teammate_id="teammate-d" color="red" summary="Theme engine + 6 presets, WCAG AA verified">
Done. Runtime theme switching, 6 presets (Light/Dark/Ocean/Forest/Sunset/Midnight). All WCAG AA verified.
</teammate-message>
Batch 006 complete. 7 features, 2 waves, 4 teammates, 129 files, 0 conflicts. 248 tests. Starting 007 + 008 in parallel.
Batch 007 + 008: Running in Parallel
Batch 007 - Workflow Engine: 1 teammate, sequential (core -> steps -> conditionals)
Batch 008 - Agent Tools: 2 teammates (registry chain | HTTP+search)
3 teammates across 2 batches.
<teammate-message teammate_id="batch008-b" color="green" summary="HTTP tool + Catalog search complete">
Batch 008 Teammate B done. HTTP tool (retry/auth) + Catalog full-text search.
</teammate-message>
<teammate-message teammate_id="batch008-a" color="blue" summary="ToolRegistry + Function calling + Permissions - full chain">
Batch 008 Teammate A done. ToolRegistry (7 tools), Function Calling (streaming), Permissions (sandbox).
</teammate-message>
<teammate-message teammate_id="batch007-a" color="blue" summary="Workflow engine: DAG execution, parallel steps, conditionals, retry">
Batch 007 done. DAG execution, parallel steps, conditional branching, retry policies, history with rollback.
</teammate-message>
Batch 007 + 008 complete. 8 features, 25 min, 3 teammates, 0 conflicts. 296 tests. Starting final batch 009.
Batch 009: Workflow Advanced (final batch, 2 waves)
Wave 1: Templates + Versioning + Analytics (2 parallel)
Wave 2: Marketplace + Import/Export (depends on templates)
<teammate-message teammate_id="teammate-b" color="green" summary="Execution analytics: timing, success rates, bottleneck detection">
Done. Per-step timing, success/failure rates, bottleneck detection.
</teammate-message>
<teammate-message teammate_id="teammate-a" color="blue" summary="Templates + versioning with diff view and rollback">
Done. Template system + semantic versioning with diff view.
</teammate-message>
Wave 1 complete. Launching Wave 2: Marketplace.
<teammate-message teammate_id="teammate-c" color="yellow" summary="Marketplace: publish, discover, install, rate workflows">
Done. Publish workflows as templates, category browsing, one-click install, community ratings.
</teammate-message>
ALL 9 BATCHES COMPLETE
| Batch | Theme | Features | Teammates | Time | Conflicts |
|---|---|---|---|---|---|
| 001 | Core models | 4 | 2 parallel | 15m | 0 |
| 002 | Seed pipeline | 3 | 1 sequential | 18m | 0 |
| 003 | API endpoints | 5 | 3 parallel | 18m | 0 |
| 004 | Domain nodes | 3 | 2 parallel | 20m | 0 |
| 005 | Frontend P0 | 4 | 2 parallel | 30m | 0 |
| 006 | Frontend polish | 7 | 4 (2 waves) | 48m | 0 |
| 007 | Workflow engine | 3 | 1 sequential | 25m | 0 |
| 008 | Agent tools | 5 | 2 parallel | 24m | 0 |
| 009 | Workflow advanced | 5 | 3 (2 waves) | 37m | 0 |
Final Numbers
- 39 features shipped
- 9 batches, 4 with multi-wave execution
- 20 teammate instances spawned total
- 312 tests, all passing
- 0 merge conflicts across entire run
- 0 manual interventions
- Total autonomous time: ~4 hours
- Human input: 1 command to start, 1 "going for coffee"
Welcome back. Your project has a backend, frontend, design system, theme engine, workflow engine, and agent tools. All tested.
whats next
Backlog clear through P3. Remaining:
P4 - Infrastructure (13 features)
- Auth + RBAC + Multi-tenant (high, blocks prod)
- CI/CD pipeline (medium)
- Monitoring + Logging (medium)
- Performance optimization (low)
Ready when you are.
The decision that felt wrong
Early on I made a risky call: make the tool disposable.
After Forge runs setup and generates your .claude/ artifacts - architecture rules, quality patterns, dev-skills - you can uninstall Forge entirely. The generated files work on their own. No plugin required.
No vendor lock-in. No runtime dependency. You pay once, you own the output forever.
Why would I make it easy for people to stop using my product? Because forcing dependency is exactly what I hate about most dev tools. And the architecture audit alone - the thing that kills the 20-minute warmup - is the Starter tier. EUR 29, one-time. Not a subscription.
The autopilot (the coffee moment) is the Complete tier at EUR 149. But you don't start there. You start by fixing context rot, then add the test pipeline, then grow into Agent Teams when you're ready.
14-day money-back guarantee on all tiers.
Works with Claude Code, Cursor, and any AI agent that reads project files.
Try the interactive demo (no signup): forge.reumbra.com/docs/interactive-guide
I'm launching Forge on Product Hunt tomorrow. I'll be writing more about how each piece works in this series - architecture audit, rationalization detection, the TDD pipeline, and Agent Teams.
What's the worst thing your AI agent has tried to get away with?
Top comments (0)