🔧
Chapter 13 · 10 min read
The Bottleneck Elimination Framework
Every 'can you help me with this?' is a bug to fix forever
Every time your agent asks you "hey, I need X to continue" — that's a bottleneck. Most people answer the question and move on. Top operators eliminate the question forever. This chapter shows you how — on every platform.
🍕 Real-life analogy
Imagine a factory where a robot arm stops 10 times a day to ask a human "which color paint?" You could answer every time (reactive). Or you could put up a sign that says "always blue for Model A, always red for Model B" and the robot arm never asks again. That's bottleneck elimination.
The 3 Types of Bottlenecks
1. Missing Context — "What is this?"
Agent doesn't know project history, preferences, or credentials.
Fix: Add it to the Knowledge Base or Tacit Knowledge layer.
2. Missing Permission — "Can I do this?"
Agent knows how but isn't authorized (e.g., tweet, deploy, send email).
Fix: Grant appropriate autonomy level (see Progressive Trust chapter) or set specific guardrails.
3. Missing Logic — "How do I decide?"
Agent faces a fork in the road and doesn't have a decision framework.
Fix: Create a Decision Protocol doc with clear if/then rules.
❌ Without Bottleneck Elimination
- ✗Agent asks 'what's the deploy target?' every single time
- ✗You answer the same 5 questions before real work starts
- ✗10+ interruptions per day for decisions the agent should know
- ✗Agent productivity plateaus — never gets truly autonomous
✅ With Bottleneck Elimination
- ✓Agent reads deployment config from knowledge base automatically
- ✓Zero repeated questions — every answer is captured once, forever
- ✓1-2 interruptions per day for genuinely novel decisions only
- ✓Agent autonomy compounds — handles more each week
The Compound Effect
Week 1
Agent stuck 10x/day
Week 2
5 bottlenecks removed → stuck 5x/day
Month 1
20 removed → stuck 1x/day
Month 2
Agent rarely needs you. You do strategy.
The "Five Whys" Technique
When your agent stops, don't just unblock it. Ask why it stopped — then eliminate that class of problem forever.
Stop:"Should I use Tailwind or CSS Modules for this component?"
Why?Because I didn't specify a preference.
Why?Because it's not in the project README.
Why?Because we never wrote a tech stack standard.
Fix:Create a standards doc: "Always Tailwind + shadcn/ui." Agent never asks again.
Bottleneck Log Template
knowledge/resources/bottleneck-log.md
# Bottleneck Log ## 2026-02-20 - **Stop:** Asked for Stripe API key - **Type:** Missing Context - **Fix:** Added key to .env + documented in project.md - **Status:** ✅ Eliminated ## 2026-02-21 - **Stop:** "Is this tweet tone okay?" - **Type:** Missing Logic (no decision framework) - **Fix:** Updated tacit.md with 5 examples of good vs bad voice - **Status:** ✅ Eliminated ## 2026-02-22 - **Stop:** "Database migration failed, what do I do?" - **Type:** Missing Context - **Fix:** Created troubleshooting.md with standard DB reset commands - **Status:** ✅ Eliminated
Decision Protocols (Advanced)
For complex decisions, give your agent a flowchart in text form:
knowledge/protocols/content-approval.md
# Content Approval Protocol When evaluating a draft tweet: 1. Is it controversial or political? → Yes → STOP. Ask me. Never post. → No → Continue. 2. Is it factually accurate? → No → Fix it. Cite sources. → Yes → Continue. 3. Does it mention a competitor? → Yes → Ensure tone is respectful. Never trash-talk. → No → Continue. 4. All checks pass? → Schedule for 9 AM ET. → Do NOT ask for confirmation.
🔌 Platform-Specific Bottleneck Patterns
Each platform has its own common bottlenecks. Here's what to pre-solve:
🐾 OpenClaw
- • Bottleneck: Agent asks "which Discord channel?" → Fix: Add channel IDs to MEMORY.md or knowledge base
- • Bottleneck: "Should I use cron or heartbeat?" → Fix: Create a decision protocol doc
- • Bottleneck: "Context too long" → Fix: Set up auto-compact rules, use isolated sessions for big tasks
🤖 Claude (API / Projects)
- • Bottleneck: Forgets project context every conversation → Fix: Use Projects feature, paste your AGENTS.md + knowledge base as project docs
- • Bottleneck: Asks "what framework are you using?" → Fix: Add tech stack to project instructions
- • Bottleneck: Can't access files → Fix: Upload key files to project, or use Claude Code CLI
💬 ChatGPT (Custom GPTs / API)
- • Bottleneck: Loses memory after conversation → Fix: Use Memory feature or Custom GPT instructions (paste knowledge base)
- • Bottleneck: "I can't access the internet" → Fix: Enable browsing in settings, or use API with function calling
- • Bottleneck: Inconsistent voice → Fix: Paste tacit.md examples in system prompt
🔗 LangChain / CrewAI / AutoGPT
- • Bottleneck: Agent loops endlessly → Fix: Set max iterations, add clear exit conditions
- • Bottleneck: Tool calls fail silently → Fix: Add error handling + fallback logic in chain definition
- • Bottleneck: Agents argue with each other → Fix: Define clear role boundaries, use sequential (not parallel) for dependent tasks
⚡ n8n / Make / Zapier
- • Bottleneck: Workflow breaks on unexpected data → Fix: Add data validation nodes before AI steps
- • Bottleneck: AI node returns wrong format → Fix: Add explicit output format in prompt + JSON schema
- • Bottleneck: Rate limits → Fix: Add delay nodes, batch processing, error retry logic
💻 Cursor / Windsurf / Cline
- • Bottleneck: "Which file should I edit?" → Fix: Add .cursorrules / AGENTS.md with project structure map
- • Bottleneck: Suggests wrong patterns → Fix: Add code examples in .cursorrules showing your preferred patterns
- • Bottleneck: Doesn't know about existing utilities → Fix: Reference your utils/helpers in project docs
🚀 The Transformation
Week 1: you unblock your agent 10 times a day (you're a babysitter).
Month 1: you unblock it once a week (you're a manager).
Month 3: it runs operations while you do strategy (you're the CEO).
The bottleneck log is the single most important practice in this entire playbook. Start one today.
Month 1: you unblock it once a week (you're a manager).
Month 3: it runs operations while you do strategy (you're the CEO).
The bottleneck log is the single most important practice in this entire playbook. Start one today.
Share this chapter
Chapter navigation
14 of 36