Run multiple AI agents simultaneously. Zero merge conflicts. Zero hidden bugs. 6 TRIZ inventive principles make parallel coding both fast AND reliable.
THE PROBLEM
2025-2026 industry research reveals a harsh truth: parallel AI agents produce code faster but create more problems than they solve.
SEE THE DIFFERENCE
Scenario: Build a user authentication feature. 5 tasks — API routes, login form, JWT middleware, database schema, and unit tests. Watch what happens.
auth.ts — API routes
auth.ts — JWT middleware ⚠️ same file!
auth.ts
auth.ts 💥
auth.ts, B→auth.ts ⚠️ overlap!
auth.ts (builds on A's work)
THE SOLUTION
TRIZ analyzed 2.5 million patents to find universal problem-solving patterns. CodyMaster applies 6 of those principles to resolve the Speed ↔ Quality contradiction.
File-dependency graph ensures only TRULY independent tasks run together. Two agents can't touch the same file. Eliminates 100% of merge conflicts before dispatch.
Each agent validates its OWN changes — syntax, tests, self-review — before reporting completion. Bugs are caught at the source, not batched at the end.
Pre-flight check scans the conflict ledger BEFORE every dispatch. If a running agent already holds a file, new tasks are blocked — never launched into danger.
Batch size starts at 2 and adapts: 3 clean runs → scale up (max 5). Any conflict → scale down (min 1 = sequential). The system learns from its own history.
Real-time conflict ledger tracks every file modification. When an agent finishes, actual files are compared against others. Overlap detected → immediate halt.
Each agent is 3 personas in one: Implementer (code) + Tester (write tests) + Reviewer (self-review). Reports DONE only when all three roles pass.
HOW IT WORKS
Extract file dependencies from task descriptions using pattern matching. Know which files each task will touch.
Build a dependency graph. Group tasks into independent batches using graph coloring — only non-overlapping tasks in the same batch.
Read parallel history. Compute optimal batch size based on past success/conflict rates. Start cautious, scale with confidence.
Check the conflict ledger against any running agents. No overlap → safe to dispatch. Any conflict → block and wait.
Send independent tasks to agents with strict Quality Contracts. Each agent knows its file scope, timeout, and required checks.
Agents report actual modified files back to the conflict ledger. Real-time detection catches unexpected overlaps instantly.
Each agent runs its mini quality gate — syntax + tests + self-review — before reporting completion. No "trust me it works."
Save session metrics to parallel history. Future batch sizes are informed by real data, not hunches.
COMPARISON
| Aspect | Normal Parallel | TRIZ-Parallel |
|---|---|---|
| Conflict Prevention | ✗ None | ✓ Dependency Graph |
| Quality Check | Once at end | ✓ Per-Agent Gate |
| Batch Sizing | Fixed (manual) | ✓ Adaptive |
| Agent Role | Code only | ✓ Code+Test+Review |
| Conflict Detection | ✗ Post-merge (late) | ✓ Real-time Ledger |
| Rollback | ✗ Manual revert | ✓ Auto-halt + Scale Down |
| Learning | ✗ None | ✓ Session History |
ROLLBACK & SAFETY
Dependency graph blocks conflicting tasks. Pre-flight scans the ledger before every dispatch.
Real-time conflict ledger catches unexpected overlaps instantly. Agents self-validate before reporting.
Adaptive batch sizing throttles down on conflict. Minimum size = 1 → sequential fallback guaranteed.
Conflict → halt → no merge → code is safe. Each agent in its own worktree via cm-git-worktrees.
Open source. Backed by TRIZ science. Install in 30 seconds.
Get Started →