bloat-detector

name	bloat-detector
description	Detect the 7 types of agentic bloat in a codebase or system design
user-invocable	true
argument-hint	["file-or-directory-path"]
allowed-tools	Read, Grep, Glob, Write
model	sonnet

Bloat Detector

Scan a codebase or system description for the 7 types of agentic bloat and produce a scored audit.

Lean Principle

• #1 Context is a liability
• #2 Right-size your model
• #3 Orchestration is not a playground
• #4 Reflections cost compute
• #5 RAG isn't always right
• #7 Reuse over recompute
• #8 More tools, more problems

Instructions

Step 1: Scan for each bloat type

Use Grep and Read to look for evidence of these 7 bloat patterns:

1. Model Bloat — Look for hardcoded frontier model names (gpt-4, claude-opus, gemini-ultra) used for simple tasks (classification, extraction, formatting).

2. Orchestration Bloat — Count distinct agent definitions, multi-agent chains, sequential agent invocations. Flag if more than 3 agents are chained sequentially.

3. Reflection Bloat — Look for self-critique patterns: "review your output", "improve your answer", retry loops, while loops around LLM calls without exit conditions.

4. Context Bloat — Look for large system prompts (>2000 chars), full document injection, unbounded conversation history loading, many tool descriptions.

5. Tool Bloat — Count tool/function definitions per agent. Flag agents with more than 5 tools.

6. Memory Bloat — Look for unbounded storage: missing TTLs, "store everything", no cleanup/expiry logic, full transcript storage.

7. Retrieval Bloat — Look for RAG pipelines that run on every query without gating: always-on vector search, no relevance thresholds, no caching of retrieval results.

Step 2: Score each category

Score each bloat type 0-3:

• 0 = No evidence of bloat
• 1 = Minor (some instances, manageable)
• 2 = Moderate (multiple instances, should optimize)
• 3 = Severe (systemic, causing significant waste)

Step 3: Write report

Write a report to reports/bloat-audit.md with:

• Score per category (0-3)
• Total score out of 21
• Specific file:line references for each finding
• Severity rating: 0-5 Lean, 6-12 Moderate bloat, 13-21 Severe bloat
• Top 3 recommended fixes

Output

Report saved to reports/bloat-audit.md

Example

Input: A project with 5 chained agents all using GPT-4, no caching, 10 tools per agent. Output:

Model Bloat: 3/3 — All 5 agents use frontier model for mixed tasks
Orchestration Bloat: 2/3 — 5 agents chained, 2 could be merged
Tool Bloat: 3/3 — 10 tools per agent, most unused
Total: 15/21 — SEVERE bloat. Start with model routing and tool pruning.