YC/YouTube/25/08/Advanced Context Engineering for Agents

• Advanced Context Engineering for Agents - YouTube

  • Video

    • • 00:00 Introduction and Background

        • Speaker: Person/Dex Horthy Role/Founder of humanlayer (YC Fall 24 batch)
        • Context Engineering History:
          • Wrote "12 Factor agents - principles of reliable LLM applications" manifesto 2025-04-22 Tue - humanlayer/GitHub/12-factor-agents
          • Term "context engineering" coined on June 4th
          • Talk titled "12 factor agents" at AI Engineer conference
        • Current Focus: Moving beyond basic context engineering to advanced techniques for coding agents

      • 00:55 Industry Insights and Challenges

        • Favorite Talks from AI/Engineer/World's Fair/25 00:55

          • Context eng: Person/Dex Horthy/YouTube/25/07/12-factor-agents

            Spec driven coding: Person/Sean Grove/YouTube/25/07/The New Code
            Dev Productivity: Person/Yegor Denisov-Blanch/YouTube/25/07/Does AI Boost Developer Productivity - 100k Devs Study - Stanford
            

        • Sean Grove's "New Code" Talk - Person/Sean Grove/YouTube/25/07/The New Code

          • Vibe Coding Problem: Sitting with agents for hours specifying requirements then discarding prompts
          • Analogy: Like Java developers committing compiled JARs instead of source code
          • Key Insight: In AI-driven development, specifications and descriptions become the important artifacts

        • Stanford Study Findings Person/Yegor Denisov-Blanch/YouTube/25/07/Does AI Boost Developer Productivity - 100k Devs Study - Stanford

          • Research Scope: 100,000 developers across enterprises and startups
          • Key Finding: AI engineering leads to significant rework
          • Limitations:
            • Doesn't work well for complex or brownfield tasks
            • Can be counterproductive for legacy codebases
            • May actually slow down development in some cases

      • 02:34 Personal Journey and Transformation

        • Forced Adoption of Spec-First Development

          • Challenge: Working with AI coder producing 20,000-line Go PRs
          • Complexity: Systems code with race conditions, shutdown order, etc.
          • Solution: Adopted specification-first development approach
          • Transformation: 8-week process from code review to spec review

        • Current Workflow

          • Success Metrics: Shipped 6 PRs in one Thursday
          • No Code Files: Haven't opened non-markdown files in editor for 2 months
          • Focus: Reading tests and specs instead of every line of code

      • 03:26 Advanced Context Engineering Goals

        • Core Objectives

          • Works well in big, complex codebases
          • Solves big, complex problems
          • No slop - production-ready code
          • Team stays aligned
          • Maximize token usage efficiently

        • Naive Approach Problems

          • Shouting back and forth until context exhaustion
          • Giving up or crying from frustration
          • Inefficient iteration cycles

      • 04:18 Intentional Compaction Strategy

        • When to Start Over

          • Signs: Agent going off track, making repeated mistakes
          • Strategy: Stop and restart with fresh context
          • Steering: Add specific guidance about what not to repeat

        • Context Management

          • Avoid: /compact command (considered ineffective)
          • Preferred: Write progress files for context transfer
          • Purpose: Onboard next agent into ongoing work

        • Context Window Optimization

          • Space Consumers: File searching, flow understanding, edits, MCP tool responses
          • Focus: Intentional compaction of progress and status

      • 05:07 Context Engineering Principles

        • LLM as Pure Functions

          • Key Insight: Only input quality improves output quality
          • Context Window: Critical for agent decision-making
          • Agent Loops: Constantly determining next tool/action

        • Optimization Priorities

          • Correctness: Ensure information accuracy
          • Completeness: Include all necessary details
          • Size: Keep context manageable
          • Trajectory: Maintain direction (vibes-based)

        • Context Quality Equation

          • Worst: Bad information
          • Second Worst: Missing information
          • Third: Too much noise
          • Target: ~170,000 tokens with efficient usage

      • 06:45 Sub-Agents and Context Control

        • Sub-Agent Strategy

          • Purpose: Context control and delegation
          • Common Use: Finding files, understanding information flow
          • Workflow: Parent agent calls sub-agent, gets focused results
          • Benefit: Parent avoids context burden of searching/reading

        • Sub-Agent Challenges

          • Telephone Problem: Information distortion through delegation
          • Prompting Complexity: How to instruct parent to instruct child
          • Deterministic vs Non-deterministic: Chaos in complex systems

      • 07:32 Frequent Intentional Compaction

        • Development Workflow Phases

          • Research: Understand system, identify relevant files, locate problems
          • Plan: Specify all changes, testing strategy, verification steps
          • Implement: Execute the plan with code generation

        • Context Management Target 09:00

          • Goal: Keep context utilization under 40%
          • Strategy: Constant plan updates and context refresh
          • Benefit: Fresh context window for each phase

        • 3 phases - Research, Plann, Implement, each with a prompt

          • Research Output: File names, line numbers, system understanding
          • Planning Output: Change specifications, testing approach, verification steps
            • "Tell me every single change you're going to make"
            • "Be very specific about how we're going to implement
          • Open Source: All prompts available publicly

      • 09:08 Implementation and Code Review

        • Workflow Benefits

          • Efficiency: Good plans eliminate need for agent shouting
          • Quality: Plans are shorter than code but catch problems early
          • Human Review: Research files easier to review than 2000-line PRs

        • Code Review Purpose

          • Mental Alignment: Keep team aware of system changes
          • Efficiency: Review 200-line plans vs 2000-line code
          • Early Problem Detection: Catch issues before implementation

      • 09:53 Real-World Validation

        • BAML Codebase Challenge

          • Scope: 300,000-line Rust codebase
          • Approach: One-shot fix attempt
          • Result: PR merged successfully
          • Validation: Works in brownfield codebases, no slop

        • Complex Problem Solving

          • Challenge: Adding WASM support to programming language
          • Time: 7 hours with Boundary CEO
          • Output: 35,000 lines of code
          • Efficiency: 1-2 weeks of work compressed to 7 hours

        • Key Insight Hierarchy

          • Bad Code Line: One bad line
          • Bad Plan: Hundreds of bad lines
          • Bad Research: Thousands of bad lines
          • Focus: Spend time on problem specification and system understanding

      • 10:50 Results and Team Transformation

        • Success Metrics

          • Token Usage: Significant monthly consumption (team of 3)
          • Efficiency: Worth the cost due to time savings
          • Team Growth: Intern shipped 2 PRs first day, 10 PRs by day 8

        • Workflow Transformation

          • Code Review: No longer reading code, only specs
          • Team Adoption: New communication and work structuring methods
          • Scalability: Working with 6-person startups to 1000-person public companies

      • 11:36 Future Outlook and Next Steps

        • Industry Trends

          • Coding Agents: Becoming commoditized
          • Team Transformation: The hard part of adoption
          • Communication: New ways of structuring work

        • Human Layer's Role

          • Support: Helping teams figure out new workflows
          • Events: Hyperengineering event (near capacity)
          • Content: 90-minute video with Vibbov discussing BAML challenge

        • Final Message

          • Urgency: Figure this out or face difficulties
          • Opportunity: Advanced context engineering for coding agents
          • Resources: QR code with all materials and links