Mar 6, 2026 · 24m · latent-space
Why Your AI Agents Don’t Work with Dex Horthy of HumanLayer | In-Context Cooking
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of In-Context Cooking, host Alan and HumanLayer CEO Dex Horthy prepare Dan Dan noodles while exploring the principles of context engineering, agentic AI reliability, and Dex's career path from NASA JPL to founding HumanLayer.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 1.3 Guest teaching 3.3 Guest disagreement 1.1 The hosts pushing back 0.0
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 19:23 Pet peeve against Twitter token flexing
Dex strongly criticizes developers who brag on social media about token consumption and parallel agent workflows instead of delivering reliable, paying software products.
Hardest push from the hosts ▶ 14:16 Questioning whether context engineering is still necessaryAlan presses Dex on whether massive model context windows will eventually render intentional context engineering obsolete.
Biggest teaching moment ▶ 13:16 Explaining model capability frontiersDex reframes model progress by explaining that skilled context engineers achieved Opus 4.5 performance levels months earlier on Opus 4.0, illustrating how prompt architecture unlocks capability.
The host holds their own ▶ 20:23 Framing enterprise AI as quality over quantityAlan synthesizes Dex's critique of vibe coding into the core reality of enterprise engineering, highlighting that mission-critical systems cannot adopt a 'move fast and break things' approach.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Tasting Benchmark Noodles and Dex's JPL NASA Internship | 1 | 3 | 0 | 0 | Alan sets up the cooking challenge and asks an open biographical question about Dex's high school JPL internship. Dex details his naive implementation of Dijkstra's algorithm for lunar crater pathfinding in an educational, friendly manner. | |
| The Coining of Context Engineering and AI Slop | 2 | 5 | 2 | 0 | Dex explains how coining context engineering came from recognizing the stark contrast between public open-source demo hype and the reliability engineering done by top 1% enterprise builders. Alan prompts for clarification on incentives and slop definitions without pushing back. | |
| Formulating the Dan Dan Sauce and Kitchen Banter | 1 | 2 | 0 | 0 | The exchange revolves around sauce preparation and Dex's career transition at Replicate from backend engineering to customer-facing sales engineering. Alan facilitates the conversation with casual prompts. | |
| HumanLayer's Focus and Scaling Coding Agents | 1 | 3 | 1 | 0 | Dex describes HumanLayer's mission to help engineering teams deploy coding agents on complex codebases via tooling and whiteboarding workshops. Alan asks about customer segmentation across startups and enterprise. | |
| Context Engineering Evolution and Frontier AI Models | 2 | 6 | 2 | 0 | Alan asks whether larger context windows diminish the need for context engineering. Dex counters by asserting that top practitioners achieve frontier results on older models, and context engineering constantly redefines the moving boundary of model capabilities. | |
| The 40% Dumb Zone and Building Context Intuition | 2 | 5 | 1 | 0 | Alan inquires about heuristics for navigating the context dumb zone. Dex clarifies that the 40% rule is training wheels for beginners, explaining that true intuition and framework selection require extensive trial and error with models. | |
| Future of Software Engineering: Coding vs Shipping | 2 | 5 | 4 | 0 | Dex takes aim at Twitter vibe coders who boast about token counts without shipping reliable software, invoking Guillermo Rauch's distinction between coding and shipping. Alan echoes the sentiment on prioritizing quality over quantity. | |
| The Final Cooking Rush and Plating Dan Dan Noodles | 1 | 1 | 0 | 0 | A collaborative cooking climax where Alan and Dex finish their Dan Dan noodles under time pressure and compare their flavor profiles and plating. | |
| HumanLayer Announcement and Judges' Final Verdict | 0 | 0 | 0 | 0 | Dex plugs HumanLayer's agentic IDE and social handles before the judge evaluates both dishes and awards Dex the win for higher meat content. |
Statements from this episode (7)
Horthy: Top 1% of AI builders don't use popular public frameworks
“The way the top one percent build is so different from the bottom 99%. You have all your, like, indie hackers and, like, open source frameworks that are very, very popular. And, like, everyone uses, and that's what you see in public, and then you go see how re…”
Horthy: AI demos work at 80% accuracy, enterprise software needs far more
“I think there's a difference between, like, people who want to build, like, Reliable software and people who want to build a cool demo. I think that's the core. The incentives are different. Yeah. The incentives are of like, okay, if this is right, 80% of the …”
Horthy: Skilled Context Engineers Matched Opus 4.5 Output on Opus 4.0
“When Opus 4.5 came out, they're like, oh, This is good enough. This is a big change, and I saw the things that they were shipping, and I was like, okay, but I know a bunch of engineers who got really good at context engineering, and they were getting the same …”
Horthy: Great AI Products Are Built at Boundary of Model Capabilities
“There will always be a thing that the model can like only kind of get right reliably. Like you find a thing that's right on the boundary of the model's capabilities, and you figure out how to get it right over and over and over again.”
Horthy: Beginners should compact LLM context at 40% of window capacity
“If you don't know what you're doing, and you don't really know what the AI model is capable of, and you don't have a lot of experience, like, you know, training wheels is like, when you get to 40%, start thinking about wrapping it up, or like doing a, like, yo…”
Horthy: Software engineers will orchestrate rather than manually write code
“I don't believe the like software engineering is dead and there'll be more, no more coders. I think the way I would describe it is like the role of the software engineer will change from like write working code to like produce working code. Or like cause worki…”