## Core Framework **Managing multiple AI coding agents simultaneously requires the "piggy farmer" mindset: patience, accountability, and realistic expectations.** Steve Yegge's real-world framework from running 5 AI agents for months reveals that agentic coding is not a panacea but a journey—requiring orchestration skills fundamentally different from traditional programming. ## The "Piggy Farmer" Metaphor From Steve Yegge's Twitter thread (February 2026): > "My piggies are the best. I love Amp. I love coding this way. I'm working via deep, ongoing conversations, keeping my piggies on track and accountable." **The Metaphor Decoded**: - **Piggies** = Individual AI coding agents (Amp, Anthropic Claude Code, etc.) - **Farmer** = Developer orchestrating multiple agents - **Farming** = Patient management through ongoing conversations and accountability - **Not**: Micromanaging every action - **But**: Setting direction, checking progress, course-correcting ## Real-World Setup Yegge's actual configuration (running for "a couple months"): 1. **Agent 1**: Working on Emacs 2. **Agent 2**: Building Node ("little piggy") 3. **Agent 3**: Porting old tests ("little piggy") 4. **Agent 4**: Code reviews 5. **Agent 5**: Porting Ruby to Kotlin ("Piggy Five") **Management approach**: "Deep, ongoing conversations, keeping my piggies on track and accountable." ## The Three Essential Principles ### 1. Patience > "Mostly, though, it's patience. The only way you can be successful with agentic coding is through patience, persistence, and low expectations." **Why patience is critical**: - Agents make mistakes frequently - Learning agent capabilities takes time - Integration requires multiple iterations - Debugging agent outputs is slow - Economic costs accumulate ($100s/week) **Anti-pattern**: Expecting agents to work perfectly from day one ### 2. Persistence **The Journey Framing**: > "It was not easy to learn how to work like this. It has been a journey full of traps and surprises. But it has paid off in droves." **What persistence means**: - Continuing despite frustrating failures - Learning each agent's strengths and failure modes - Building institutional knowledge about orchestration - Not giving up when agents produce garbage - Treating it as skill development, not tool adoption ### 3. Low Expectations > "Coding agents aren't a panacea. They can't even be trusted. You have to be very patient with them." **Reality check**: - ❌ "Agents will solve everything automatically" - ✅ "Agents are unreliable assistants that occasionally produce gold" - ❌ "Set and forget" - ✅ "Ongoing supervision and course-correction" - ❌ "They understand context perfectly" - ✅ "They misunderstand frequently and require clarification" ## Economic Reality > "At least, it has for Anthropic, who bilk me, sorry I mean 'bill me', for a few hundred bucks a week." **Cost structure**: - Multiple agents running simultaneously - Hundreds of dollars weekly in API costs - Significant investment for individual developers - ROI comes from productivity gains over months, not days **Implication**: Multi-agent orchestration is expensive. You're paying for parallelization and exploration, not perfect execution. ## The "Secret Sauce" Myth Yegge addresses the common question about his approach: > "There are naysayers. I can share with you the secret sauce they are all missing. Is it money, you ask? No, sure. None of us have enough money for this shit." **The real secret**: It's not money. It's not special techniques. It's **patience, persistence, and low expectations**. ## Management Approach: "Deep, Ongoing Conversations" **Not**: Fire-and-forget task assignment **But**: Continuous dialogue and accountability ### Conversation Pattern 1. **Assign task** to specific agent 2. **Check progress** regularly 3. **Course-correct** when agent misunderstands 4. **Keep accountable** - agents need reminders 5. **Manage expectations** - celebrate small wins ### Key Insight The "piggy farmer" doesn't micromanage individual actions but maintains **ongoing relationships** with each agent through conversation. ## The Payoff: "Paid Off in Droves" > "But it has paid off in droves. At least, it has for Anthropic, who bilk me..." **What "paid off" means**: - Significant productivity increase (despite costs) - Parallel work on 5 different codebases/tasks - Learning journey that builds orchestration skills - New way of working that feels superior despite challenges **The joke**: Most immediate beneficiary is Anthropic (collecting API fees), but the developer gains productivity that justifies the cost. ## Comparison to Other Frameworks ### vs. Factory Farming Code - **Factory Farming**: Industrial-scale code production - **Piggy Farming**: Managing specific agents with individual personalities/capabilities - **Connection**: Both involve treating AI code generation as agricultural-scale operation ### vs. Pair Programming - **Pair Programming**: Synchronous, single agent, interactive - **Piggy Farming**: Asynchronous, multiple agents, supervisory - **Shift**: From collaborative partner to orchestrator of workers ### vs. The Merge Wall - **Merge Wall**: Bottleneck of integrating parallel agent outputs - **Piggy Farming**: Addresses the management side (keeping agents productive) - **Connection**: Multiple parallel agents create merge challenges ## Cross-Domain Applications ### Software Development Teams - **Pattern**: Managing multiple junior developers - **Parallel**: Each "piggy" like a junior developer needing guidance - **Skill**: Delegation with accountability, not micromanagement ### Project Management - **Pattern**: Managing contractors or outsourced teams - **Parallel**: Agents are like contractors—you don't control their process, only outcomes - **Skill**: Clear task definition and outcome verification ### Personal Productivity - **Pattern**: Managing multiple parallel projects - **Parallel**: Context-switching between different "agents" (projects) - **Skill**: Maintaining momentum across multiple workstreams ### AI Strategy - **Pattern**: Multi-model deployment for different tasks - **Parallel**: Different LLMs excel at different tasks (like different agents) - **Skill**: Model selection and task routing ## Practical Implementation ### Starting Small 1. **Two agents first**: Don't jump to 5 immediately 2. **Different tasks**: Ensure agents aren't competing for same codebase 3. **Daily check-ins**: Review each agent's progress daily 4. **Low stakes**: Start with non-critical tasks ### Scaling Up 1. **Add agents gradually**: Once comfortable with 2, add 3rd 2. **Task partitioning**: Clear boundaries prevent conflicts 3. **Conversation rhythm**: Develop check-in cadence for each agent 4. **Cost monitoring**: Track API spending as agents multiply ### Warning Signs - **Agent conflict**: Multiple agents editing same files - **Runaway costs**: Bills exceeding productivity value - **Overwhelming complexity**: Too many agents to track - **No progress**: Agents spinning without meaningful output ## The "Piggy Farmer" Mindset ### What It Requires - **Comfortable with ambiguity**: Agents will surprise you - **Patient with failure**: Most attempts fail initially - **Persistent despite costs**: Economic investment required - **Realistic about limitations**: Agents aren't magic - **Conversational management style**: Talk to your piggies ### What It Produces - **Parallelized productivity**: 5 tasks advancing simultaneously - **Orchestration skill**: New capability beyond traditional coding - **Cost-effective (eventually)**: ROI positive after learning curve - **Competitive advantage**: "Piggy farmers" outpace traditional developers ## Related Concepts - [[Pair Programming to Parallel Delegation Shift]] - Mental model transformation from synchronous to asynchronous AI collaboration - [[Factory Farming Code]] - Industrial-scale code production paradigm - [[The Merge Wall]] - Integration bottleneck from parallel agent work - [[2000-Hour AI Trust Curve]] - Learning to predict agent behavior over time - [[LLM Mechanical Sympathy]] - Understanding agent capabilities through operation - [[AI Productivity Paradox]] - Initial productivity drop before gains - [[Agency Preservation Standard]] - Maintaining human understanding despite delegation - [[Software Development Methodology]] - Adapting development process for AI agents ## Common Obstacles ### "I can't afford hundreds per week" **Reality**: Start with fewer agents, use cheaper models, or batch work to reduce costs. The framework scales down—you don't need 5 agents to benefit from the mindset. ### "My agents keep failing" **Expected**: That's why patience is essential. Failure is the norm, not the exception. The skill is learning which failures to tolerate and which to fix. ### "This seems inefficient" **Paradox**: It IS inefficient short-term. The efficiency comes from parallel exploration and long-term learning, not immediate perfect execution. ### "How do I know which agent to use for what?" **Learning curve**: [[2000-Hour AI Trust Curve]]—it takes time to learn agent strengths. Start with obvious task partitioning (backend vs frontend, different languages, etc.) ## Steve Yegge Context **Who**: Former Amazon and Google engineer, known for influential tech blog posts **Credibility**: Decades of software engineering experience, early adopter of AI tooling **Perspective**: Pragmatic skeptic who found success despite initial difficulties **Message**: This works, but it's hard—don't expect magic ## Source Extracted from [[Steve Yegge on Using Multiple Coding Agents]]—Twitter thread describing his experience running 5 Amp/Code agents simultaneously for months. **Screenshots**: Twitter posts showing real-world multi-agent orchestration and the "piggy farmer" metaphor. ## Verification [Verified] - Content synthesized from Steve Yegge's actual Twitter thread with direct quotes about multi-agent orchestration experience, costs, challenges, and payoffs.