1 · The Problem
A single AI agent writing code hits three walls:
Context Loss
When a large task is split into subtasks, each subtask loses context about the whole. Result: subagents guess imports, miss dependencies, generate incompatible code. Observed: 138 compilation errors from incorrect import guessing (Sumud, 2026-04-08)
When a large task is split into subtasks, each subtask loses context about the whole. Result: subagents guess imports, miss dependencies, generate incompatible code. Observed: 138 compilation errors from incorrect import guessing (Sumud, 2026-04-08)
Error Accumulation
Writing 5 files then compiling means errors multiply. 1 wrong import in file 1 cascades to 138 errors across 5 files. Fixing takes 30 minutes instead of 1. Root cause: no intermediate verification between file generations
Writing 5 files then compiling means errors multiply. 1 wrong import in file 1 cascades to 138 errors across 5 files. Fixing takes 30 minutes instead of 1. Root cause: no intermediate verification between file generations
Role Confusion
A single agent tries to architect, implement, test, review, and deploy simultaneously. It optimizes for the last instruction, not the holistic quality. Ref: Qian et al. (2023) — role separation reduces hallucination by 66%
A single agent tries to architect, implement, test, review, and deploy simultaneously. It optimizes for the last instruction, not the holistic quality. Ref: Qian et al. (2023) — role separation reduces hallucination by 66%
2 · The Solution: Three Layers
LAYER 3: PRODUCTION — Agents that OPERATE the product ♚ Rey: Content strategy, campaigns, crisis response ♛ Reina: Ethical gate, content quality review ♜ Torre: Publishing to social media channels ♝ Alfil: Community engagement, moderation ♞ Caballo: Sentiment analysis, anomaly detection ♟ Peon: Data sync, scheduled maintenance LAYER 2: DEVELOPMENT — Agents that BUILD the product ♚ Rey: Architecture decisions, task decomposition ♛ Reina: Code review, security gate ♜ Torre: Backend implementation ♝ Alfil: Frontend implementation ♞ Caballo: Testing (unit, integration, E2E) ♟ Peon: Build, deploy, CI/CD LAYER 1: INFRASTRUCTURE — Shared tools and communication MCP Server | Blackboard | Git | Database | Compiler | Test Runner
Key: Each layer uses the same 6 chess pieces but for different domains. Development agents build the code. Production agents run the product. Infrastructure is shared.
3 · The Agents
| Piece | Role | Can Do | Cannot Do | Cognitive Model |
|---|---|---|---|---|
| ♚ Rey | Architect | Read code, write decisions, decompose tasks | Write code, deploy, commit | System 2: slow, deliberate (Kahneman) |
| ♛ Reina | Reviewer | Read code, review diffs, run tests (read-only) | Write code, deploy | Dual-process: fast heuristics + slow analysis |
| ♜ Torre | Backend | Read/write backend code, run compiler | Deploy, merge to main, modify CI | Pattern-matching for CRUD, deep for novel logic |
| ♝ Alfil | Frontend | Read/write UI code, run type checker | Deploy, modify backend | Visual + structural (components, layouts) |
| ♞ Caballo | Tester | Read code, write tests, run test suite | Write production code | Adversarial: finds bugs, not confirms success |
| ♟ Peon | DevOps | Run deploy scripts, manage infra | Write application code | Procedural: follows protocols precisely |
Principle of Minimal Authority (Saltzer &
Schroeder, 1975): Each agent has the minimum capabilities for its
role. Torre cannot deploy. Caballo cannot write production code.
This prevents entire classes of errors.
4 · Communication: The Blackboard
Agents don't talk to each other directly. They read and write to a shared blackboard. This prevents context loss — the #1 problem in multi-agent systems.
┌──────────────────────────────────────────────┐ │ BLACKBOARD │ │ │ │ context/ │ │ ├── current_task.yaml ← Rey writes │ │ ├── stack_profile.yaml ← Rey writes │ │ └── import_map.json ← Torre writes │ │ │ │ signals/ │ │ ├── compile_result.json ← Compiler writes │ │ ├── test_result.json ← Test runner writes│ │ └── lint_result.json ← Linter writes │ │ │ │ artifacts/ │ │ └── files_in_progress/ ← Torre/Alfil write │ │ │ │ reviews/ │ │ └── pending/ ← Reina writes │ │ │ │ ALL AGENTS CAN READ EVERYTHING │ └──────────────────────────────────────────────┘
Import Map Protocol (prevents the 138-error problem)
Before splitting any large file, Torre writes an import map to the blackboard. Every submodule reads this map to get correct imports. No guessing.
BEFORE refactoring traveler_bot.rs (3675 lines):
1. Torre reads entire file
2. Torre extracts: imports, public functions, cross-references
3. Torre writes import_map.json to blackboard
4. For each submodule:
a. Read import_map from blackboard
b. Write ONE file
c. cargo check ← CRITICAL: verify IMMEDIATELY
d. Fix any errors
e. Proceed to next file 5 · Task Lifecycle
┌──────────┐ ┌──────────┐ ┌──────────┐
│ CREATED │────▶│ ACTIVE │────▶│ TESTING │
│ (Rey) │ │ (Torre/ │ │ (Caballo)│
└──────────┘ │ Alfil) │ └──────────┘
└──────────┘ │
Each file: ▼
write → compile → fix ┌──────────┐
write → compile → fix │ REVIEW │
write → compile → fix │ (Reina) │
└─────┬────┘
Pass? │
┌────Yes───┴───No────┐
▼ ▼
┌────────┐ ┌──────────┐
│DEPLOY │ │ RETURNED │
│ (Peon) │ │ (fix it) │
└────────┘ └──────────┘ Contracts: Pre and Post Conditions
Every task has machine-verifiable conditions. If postconditions fail, the task is not done.
Preconditions (before starting)
- Code compiles
- All tests pass
- File exists / function exists
Postconditions (must be true when done)
- Code compiles (mandatory)
- All tests pass (mandatory)
- New code has tests
- No file exceeds 1500 LOC
6 · The Core Loop: Generate-Check-Fix
THE #1 RULE: Compile after EVERY file. Never
write multiple files before checking. The cost of checking early
is 5 seconds. The cost of 138 accumulated errors is 30+ minutes.
Torre's work loop:
┌─────────────────────────┐
│ 1. Read context + │
│ blackboard │
└───────────┬─────────────┘
▼
┌─────────────────────────┐
│ 2. Write/edit ONE file │◀──────────┐
└───────────┬─────────────┘ │
▼ │
┌─────────────────────────┐ │
│ 3. cargo check │ ┌─────┴─────┐
│ (2-5 seconds) │ │ 4. Fix │
└───────────┬─────────────┘ │ errors │
▼ └─────┬─────┘
┌────┴────┐ │
│ Errors? ├── Yes ─────────────┘
└────┬────┘
│ No
▼
┌─────────────────────────┐
│ 5. Signal: file ready │
│ Proceed to next file │
└─────────────────────────┘ 7 · Quality Framework
Architectural Invariants (never violated)
| ID | Rule | Why |
|---|---|---|
| INV-1 | No file > 1500 LOC | Large files = context loss for agents and humans |
| INV-2 | No function > 80 LOC | Long functions = untestable, unreadable |
| INV-3 | SQL parameterized only | Prevents SQL injection (OWASP #1) |
| INV-4 | Auth on all admin endpoints | Prevents unauthorized access |
| INV-5 | Every endpoint has a test | No test = no confidence = no deploy |
| INV-6 | No secrets in code | Env vars only — secrets rotate |
| INV-7 | No auto-publish without approval | Humans approve public-facing content |
Health Score
Score = weighted average of 6 metrics (0-100): Files under LOC limit (25%) ████████████████████████░ 92 Endpoint test coverage (25%) ███████████████████░░░░░░ 76 Compiler warnings (15%) ██████████████░░░░░░░░░░░ 56 Dead code items (10%) ██████████████████░░░░░░░ 72 Duplicated patterns (15%) ██████████░░░░░░░░░░░░░░░ 40 Crate balance (10%) ███████░░░░░░░░░░░░░░░░░░ 32 OVERALL HEALTH: 66/100 — needs improvement TARGET: 80+
Test Pyramid
/ E2E \ 5-10% of tests
/ (slow, \ Playwright / Cypress
/ high value) \ Critical user journeys
/─────────────────\
/ Integration \ 60-70% of tests
/ (medium speed) \ Every API endpoint
/───────────────────────\
/ Unit Tests \ 20-30% of tests
/ (fast, isolated) \ Pure functions
/─────────────────────────────\ 8 · Cross-Layer Interactions
Production finds a bug → Development fixes it
LAYER 3 (Production):
♞ Caballo detects: "Instagram sync failing for 2 hours"
│
▼ SIGNAL: incident/sync-failure.yaml
───────────────────────────────────────────────
LAYER 2 (Development):
♚ Rey reads incident → creates task with contract
│
▼
♜ Torre implements fix → cargo check → cargo test
│
▼
♞ Caballo writes regression test
│
▼
♛ Reina reviews → approved
│
▼
♟ Peon deploys
│
▼ SIGNAL: fix deployed
───────────────────────────────────────────────
LAYER 3 (Production):
♞ Caballo verifies: sync restored ✓ Development enhances Production agents
Dev-Rey decides: "Soviet Torre needs Instagram Reels support"
│
▼
Dev-Torre implements: publish_reel() in broadcaster.rs
Dev-Caballo writes: integration test for Reels endpoint
Dev-Reina reviews: security ✓, tests ✓, quality ✓
♟ Dev-Peon deploys new binary
│
▼
Prod-Torre now has Reels capability
Prod-Rey updates content strategy to include Reels 9 · How to Start
For an existing project:
Step 1: Copy the Sumudfiles/ directory to your project root.
Step 2: Open Claude Code and say:
"Read Sumudfiles/framework/adoption.md. We're using Soviet Chess. As Rey, audit this codebase and give me a health score."
Step 3: Follow the priorities Rey identifies.
Step 4: At the end of each session, generate a postmortem.
Step 2: Open Claude Code and say:
"Read Sumudfiles/framework/adoption.md. We're using Soviet Chess. As Rey, audit this codebase and give me a health score."
Step 3: Follow the priorities Rey identifies.
Step 4: At the end of each session, generate a postmortem.
For a new project:
Step 1: Copy Sumudfiles/ to your project root.
Step 2: Fill in Sumudfiles/templates/stack-profile.yaml with your stack.
Step 3: Open Claude Code and say:
"Using Soviet Chess Framework, as Rey design the initial architecture for: [describe your project]. Create skeleton with proper module boundaries."
Step 2: Fill in Sumudfiles/templates/stack-profile.yaml with your stack.
Step 3: Open Claude Code and say:
"Using Soviet Chess Framework, as Rey design the initial architecture for: [describe your project]. Create skeleton with proper module boundaries."
Incremental adoption (for skeptics):
| Week | Adopt | Agent | Effort |
|---|---|---|---|
| 1 | Compile after every file edit | Torre | Zero — just discipline |
| 2 | Write 1 test per new endpoint | Caballo | 5 min per endpoint |
| 3 | Run security checklist before commit | Reina | 2 min per commit |
| 4 | Split any file > 1500 LOC | Rey | 1-2 hours per file |
| 5+ | Full framework with contracts | All | Built into workflow |
10 · References
- Qian, C. et al. (2023). Communicative Agents for Software Development. arXiv:2307.07924
- Hong, S. et al. (2023). MetaGPT: Meta Programming for Multi-Agent Collaborative Framework. arXiv:2308.00352
- Park, J.S. et al. (2023). Generative Agents: Interactive Simulacra of Human Behavior. arXiv:2304.03442
- Hayes-Roth, B. (1985). A Blackboard Architecture for Control. Artificial Intelligence, 26(3).
- Minsky, M. (1986). The Society of Mind. Simon & Schuster.
- Brooks, R.A. (1986). A Robust Layered Control System for a Mobile Robot. IEEE J. Robotics.
- Meyer, B. (1992). Applying Design by Contract. IEEE Computer, 25(10).
- Smith, R.G. (1980). The Contract Net Protocol. IEEE Trans. Computers.
- Beck, K. (2003). Test-Driven Development. Addison-Wesley.
- Fagan, M.E. (1976). Design and Code Inspections. IBM Systems Journal, 15(3).
- Kahneman, D. (2011). Thinking, Fast and Slow. FSG.
- Beyer, B. et al. (2016). Site Reliability Engineering. O'Reilly.
- Forsgren, N. et al. (2018). Accelerate: The Science of DevOps. IT Revolution.
- Cohn, M. (2009). Succeeding with Agile. Addison-Wesley.
- Kazman, R. et al. (2000). ATAM: Architecture Evaluation. CMU/SEI.
- Saltzer, J.H. & Schroeder, M.D. (1975). Protection of Information. Proc. IEEE.
- Claessen, K. & Hughes, J. (2000). QuickCheck. ICFP.
- Letouzey, J.P. (2012). The SQALE Method. IEEE MTD Workshop.
- Martin, R.C. (2003). Agile Software Development. Pearson.
- Shore, J. & Warden, S. (2007). The Art of Agile Development. O'Reilly.
Soviet Chess Framework v2.0 — A Multi-Agent System for Software Development. Created by David Pineda with Claude Opus 4.6 — April 2026. Licensed under CC BY-SA 4.0.