9:31am - 9:50am | No Vibes Allowed: Solving Hard Problems in Complex Codebases
Speaker: Dex Horthy, CEO, HumanLayer
Speaker Profile: Full Speaker Profile
Bio: CEO, HumanLayer
Topic: Solving hard problems in complex production codebases where AI tools struggle

Links
https://github.com/humanlayer/humanlayer
@dexhorthy
Notes
- 12 factor agents in June
- Context engineering
- Does AI agents actually help
- 100K developers across all company sizes
- a lot of rework, working more just fixing the slop from last week
- how can we get the most out of data models
- we were shipping so much we had to change the way that we collaborate
- took 8 weeks
- research plan prompt system
- refresh context
- how do you know when it s time to start of
- “you’re absolutely right” — time to start over
- intentional compaction -> compress the context down to a markdown file
- what exactly, are we compaction
- what takes up space in your context window
- correctness
- completeness
- size
- trajectory
- LLMs are stateless function
- put better tokens in to get the right tokens out
- geoff huntly did a research
- “dumb zone”
- around the 40% you are going to see diminishing turn
- use subagents
- they are for controlling context
- go find how this works (subagents)
- they are for controlling context
- “build your entire plan around context workflow”
- “keep the context under 40%”
- DO NOT OUTSOURCE THE THINKING
- it can only amplify the thinking that you’ve done
- spec driven development is broken
- what does it mean
- semantic diffusion
- the code is like assembling now and just focus o the markdown
- “memento is the best movie on context engineering”
- “on demand compressed context”
- planning == leverage
- code review -> mental alignment
- this talk had twice as much slides as the time allows
- hardness engineering as a subset of ctxe
- hard part will be term work flow and update sdlc
- hiring to build an idea
Slides
Slide: 2025-11-21-09-35

Key Point: Demonstrates the significant adoption and growth of their open-source prompt library, using GitHub stars as a metric to show thousands of users have adopted their OSS prompts over approximately one year.
Literal Content:
- Left side: “Star History” graph showing humanlayer/humanlayer GitHub stars growth from October 2023 to October (current), showing exponential growth from ~1K to 7K stars
- Right side (on black background): “1000s of folks are using our OSS prompts”
- Footer: “no vibes allowed | @dexhorthy”
- Code Summit branding
Slide: 2025-11-21-09-37

Key Point: Appears to be a title or transition slide emphasizing their data-driven, evidence-based approach (“no vibes allowed”) - likely reinforcing the theme that their work is based on measurable results rather than hype or speculation.
Literal Content:
- Large text: “no vibes allowed | @dexhorthy”
- Code Summit branding
Slide: 2025-11-21-09-50

Key Point: Illustrates a structured, human-in-the-loop development process with three distinct phases (Research, Planning, Implementation), each requiring human review and incorporating revision cycles - emphasizing the importance of human oversight in AI-assisted development rather than fully autonomous coding.
Literal Content:
- Three parallel workflow diagrams showing:
- Phase 1 Research: Research → Research Document → Human Review (with Revisions feedback loop)
- Phase 2 Planning: Planning → Implementation Plan → Human Review (with Revisions feedback loop)
- Phase 3 Impl: Impl → Clean Code → Human Review (with Revisions feedback loop)
- Author attribution: “JAKE NATIONS, SEP 04, 2025”
- Reference: “https://www.arthropod.software/p/vibe-coding-our-way-to-disaster”
- Footer: “no vibes allowed | @dexhorthy”