Modularity

Why focused components beat monolithic approaches.

The Problem

One giant model trying to do everything:

• Good at nothing in particular
• Hard to maintain
• Inflexible
• Expensive

Research

"Agentic Proposing" (arXiv:2602.03279)

Finding: 4B proposer + modular skills = 91.6% on AIME25 (math competition).

Architecture	Size	Accuracy
Monolithic (70B)	70B	78%
Modular (4B + skills)	4B	91.6%

Note: Results from AIME25 math benchmark. Demonstrates that modular approaches can outperform larger monolithic models.

Why Modularity Wins:

1. Specialization — Each component optimized for its domain
2. Composability — Mix and match as needed
3. Maintainability — Update one without affecting others
4. Efficiency — Only load what's needed

arXiv

Tachikoma's Modular Architecture

Skills

Each skill is focused, self-contained:

skills/
├── code-agent/           # Coding
├── analysis-agent/       # Review
├── research-agent/       # Investigation
├── git-commit/           # Git
├── verifier-code-agent/  # Verification

Skill Chains

Compose skills for complex workflows:

implement-verify:
  skills:
    - code-agent
    - verifier-code-agent
    - formatter

After execution: Reflect on approach, flag issues, suggest improvements.

Context Modules

Modular project-specific rules:

context-modules/
├── 00-core-contract.md
├── 10-coding-standards.md
├── 12-commenting-rules.md
├── 20-git-workflow.md
└── 30-research-methods.md

Comparison

Aspect	Monolithic	Modular
Size	3300+ lines	1550 lines
Testing	Hard	Independent
Maintenance	Coupled	Isolated
Loading	Everything	What's needed

See Also

• Skill Chains
• Skill Execution
• Research Overview