When multiple AI agents build something together, they fail. One writes processPayment(), another expects handlePayment(). Nothing fits.
Forge gives every agent a shared contract before they start, so their code works together on the first try.
flowchart LR
A["1. Analyze\n_______\nRead project,\nscore complexity"] --> B["2. Pick Mode\n_______\nParallel, Hybrid,\nor Sequential"]
B --> C["3. Contract\n_______\nShared blueprint:\nnames, types, style"]
C --> D["4. Dashboard\n_______\nLive monitor,\nzero dependencies"]
D --> E["5. Execute\n_______\nAgents build\ntheir parts"]
E --> F["6. Validate\n_______\nAuto-check\nsyntax & imports"]
F --> G["7. Deliver\n_______\nMerge into\nclean code"]
style A fill:#4a1942,stroke:#9b59b6,color:#fff
style B fill:#1a3a5c,stroke:#3498db,color:#fff
style C fill:#1a4731,stroke:#27ae60,color:#fff
style D fill:#4a3728,stroke:#e67e22,color:#fff
style E fill:#4a1942,stroke:#9b59b6,color:#fff
style F fill:#1a3a5c,stroke:#3498db,color:#fff
style G fill:#1a4731,stroke:#27ae60,color:#fff
Forge picks the right one based on how connected your modules are.
flowchart TB
subgraph PARALLEL["Parallel Mode | Coupling < 0.4 | 5.2x faster"]
direction LR
PA1[Agent 1] ~~~ PA2[Agent 2] ~~~ PA3[Agent 3] ~~~ PA4[Agent 4]
end
subgraph HYBRID["Hybrid Mode | Coupling 0.4-0.7 | 2.8x faster"]
direction LR
HCore[Core Agent] --> HF1[Feature 1]
HCore --> HF2[Feature 2]
HCore --> HF3[Feature 3]
end
subgraph SEQUENTIAL["Sequential Mode | Coupling > 0.7 | Safest"]
direction LR
S1[Agent 1] --> S2[Agent 2] --> S3[Agent 3] --> S4[Agent 4]
end
style PARALLEL fill:#1a4731,stroke:#27ae60,color:#fff
style HYBRID fill:#4a3728,stroke:#e67e22,color:#fff
style SEQUENTIAL fill:#4a1942,stroke:#9b59b6,color:#fff
Tested across 400+ experiments.
xychart-beta
title "Quality Score by Approach"
x-axis ["Naive\nParallel", "Naive\nSequential", "Shared\nContext", "Self\nRepair", "Human\nContract", "LLM\nContract", "Forge"]
y-axis "Quality Score" 0.4 --> 1.0
bar [0.571, 0.614, 0.682, 0.728, 0.794, 0.803, 0.871]
| What we measured | Result |
|---|---|
| Quality improvement | +52.5% (0.571 to 0.871) |
| Speed in parallel mode | 5.2x faster |
| Errors caught automatically | 97% (vs 58% manual) |
| Scaling to 16 agents | 8% error increase (vs 340% without contracts) |
The contract is the key idea. Before any agent writes code, Forge generates this:
flowchart TB
CONTRACT["Coordination Contract"]
CONTRACT --> MM["Module Manifest\n_What each agent builds_"]
CONTRACT --> ID["Interface Definitions\n_Function signatures & types_"]
CONTRACT --> ST["Shared Types\n_Common data structures_"]
CONTRACT --> SG["Style Guide\n_Naming, formatting rules_"]
CONTRACT --> DM["Dependency Map\n_Who imports from where_"]
CONTRACT --> SB["Section Boundaries\n_Clear code ownership_"]
style CONTRACT fill:#4a1942,stroke:#9b59b6,color:#fff
style MM fill:#1a3a5c,stroke:#3498db,color:#fff
style ID fill:#1a3a5c,stroke:#3498db,color:#fff
style ST fill:#1a3a5c,stroke:#3498db,color:#fff
style SG fill:#1a3a5c,stroke:#3498db,color:#fff
style DM fill:#1a3a5c,stroke:#3498db,color:#fff
style SB fill:#1a3a5c,stroke:#3498db,color:#fff
forge build project.yamlThat's it. Forge handles everything.
| Document | What's in it |
|---|---|
| SKILL.md | Full protocol spec |
| Architecture | How the 7 phases work |
| Experiments | All 400+ experiment data |
| Agent Roles | 15+ agent types |
| REST API Example | Worked example |
MIT
HappyCapy Research