purpose: Complete system overview of 12 Leverage Points in Agentic Coding type: foundational-knowledge domain: agentic-coding phase: phase-2-sdlc audience: LLM-optimized last-updated: 2025-09-30 The 12 Leverage Points represent systematic intervention points in Phase 2 Software Development Lifecycle where developers can dramatically amplify agent performance. These leverage points transform traditional coding from manual implementation to orchestrated agent collaboration through strategic control points. Phase 2 Software Development Lifecycle The paradigm shift where software development transitions from human-written code to AI-agent-orchestrated code generation. Humans become conductors, agents become implementers. 1950s-2023 Human writes code → Machine executes Human typing speed and cognitive load 2023-present Human orchestrates → Agent writes → Machine executes Agent context, communication, and tooling 12 Leverage Points address these bottlenecks Adopt Your Agent's Perspective critical mindset-shift Your agent is brilliant but blind. It possesses vast knowledge but zero context about your specific codebase, requirements, or environment until you provide it. Every interaction starts from nothing. No persistent memory between sessions Must rebuild context every conversation Cannot see files unless explicitly shown Must provide relevant code/docs proactively No knowledge of your tools, versions, setup Must specify environment details Knowledge cutoff date limits awareness May not know latest APIs, libraries, patterns "My agent should figure this out" "What does my agent need to see to succeed?" What context is missing for this task? Which files/docs would help my agent understand? What assumptions am I making that my agent doesn't know? How can I make this task easier for my agent? What tools does my agent need to complete this? Direct control within agent system Indirect influence through environment Direct manipulation of agent's operational parameters. These are immediate, high-impact interventions. CONTEXT What can your agent see? Determines agent's situational awareness Files shown, conversation history, documentation provided MODEL Which LLM powers your agent? Determines reasoning capability and token limits GPT-4, Claude Sonnet, Gemini, etc. PROMPT How are you communicating? Determines agent understanding and response quality Instructions, examples, constraints, tone TOOLS What can your agent execute? Determines agent's action capabilities File operations, shell commands, API calls, MCP servers Indirect optimization of agent environment. These compound over time and scale across multiple agents/tasks. DOCUMENTATION What agent-specific context exists? Persistent knowledge base for agent consumption CLAUDE.md, UFC context files, project READMEs TYPES Is information structured? Agent can validate and catch errors automatically TypeScript, Zod schemas, database schemas ARCHITECTURE Can agent navigate your code? Agent understands system boundaries and dependencies Modular structure, clear naming, separation of concerns TESTS Can agent validate its own work? Agent self-corrects without human intervention Unit tests, integration tests, validation scripts PLANNING Does agent organize work systematically? Reduces hallucination and improves consistency TodoWrite, task decomposition, memory checkpoints AI DEVELOPER WORKFLOWS (ADWs) Are repeatable patterns captured? One-shot success for common scenarios PR creation, debugging, feature implementation flows AGENTIC CODING KPIs How do you measure agent effectiveness? Data-driven optimization decisions First-attempt success rate, context utilization, hallucination rate ONE-SHOT SUCCESS Can agent succeed without iteration? Maximum velocity, minimum frustration All 11 previous leverage points optimized Direct manipulation of agent's runtime environment Immediate effect on current task Session-specific, must be repeated High granular control Visible in conversation or config Tactical adjustments during active development Shaping agent's external environment for sustained improvement Compounding benefits over time Persistent across sessions Scales to multiple agents/developers Embedded in codebase/tooling Strategic investments for long-term velocity In-Agent points provide immediate tactical wins. Through-Agent points create strategic infrastructure. Maximum effectiveness requires optimizing both simultaneously. The 12 Leverage Points are not isolated tactics but an interconnected system. Optimizing one point amplifies the effectiveness of others. Implementing a new feature Documentation: CLAUDE.md provides agent project context Context: Agent sees relevant files automatically Architecture: Agent knows where to add feature Planning: Agent breaks down implementation steps Prompt: Clear requirements enable focused work Tools: Agent writes files and runs tests Tests: Agent validates implementation One-Shot: Feature complete in single attempt Without leverage points: 5-10 iterations, 2 hours With leverage points: 1 iteration, 15 minutes Start with Core Four (1-4): Immediate impact, minimal investment Add Documentation (5), Tests (8), Planning (9): Force multipliers Integrate Types (6), Architecture (7), ADWs (10), KPIs (11), One-Shot (12): Mastery Progressive optimization beats premature perfection. Master one leverage point before adding the next. Measure impact to guide prioritization. Agents are not autonomous - they require orchestration Your bottleneck shifted from typing to context provision Small improvements in leverage points = exponential velocity gains Agent perspective is the meta-skill that unlocks all 12 points Phase 2 SDLC is not about replacing developers, but amplifying them Study Core Four In-Agent points (core-four-in-agent.md) Learn Through-Agent points (through-agent-points.md) Adopt Agent Perspective mindset (agent-perspective-mindset.md) Master Elite Context engineering (lesson-09-elite-context.md)