# Coding agents on real codebases

A pack of 6 talks from the AI Engineer YouTube channel, in the order to watch them. 1h 44m of video.
Page: https://aietalks.com/packs/coding-agents-on-real-codebases

Coding agents do fine in a fresh repo and make a mess of a ten year old one. Everyone in this pack hit that wall on real code and came back with a different reason for it. The context window fills up with junk. Nobody is tending the design between reruns. The engineers who would catch the mistakes are the ones quietly refusing to use the thing. The order matters. Horthy and Pocock diagnose the problem from two directions, Zhang and Murag supply the mechanism both of them reach for, Liguori and Blum are about what happens when a whole team does this at once, and Pant is for the code you cannot afford to be wrong.

## This pack is for you if

- Your agent's pull requests need more rework than they save.
- Senior engineers on your team have quietly stopped using it.
- You tried regenerating code from a spec and got a worse system each time.

## The talks, in order

### 1. No Vibes Allowed: Solving Hard Problems in Complex Codebases

Dex Horthy, HumanLayer | 20:31 | AI Engineer CODE 2025
Video: https://www.youtube.com/watch?v=rmvDxxNubIg
Summary: https://aietalks.com/talks/no-vibes-allowed-solving-hard-problems-in-complex-codebases.md

Why first: it names the mechanism everything else in this pack works around. Horthy's claim is that the context window is the only thing deciding the agent's next move, that quality drops once it is about 40 percent full, and that research, plan and implement exists to keep it under that line. Read this and the other five talks stop sounding like unrelated advice.

### 2. Software Fundamentals Matter More Than Ever

Matt Pocock | 18:26 | AI Engineer Europe 2026
Video: https://www.youtube.com/watch?v=v4F1gFy-hqg
Summary: https://aietalks.com/talks/software-fundamentals-matter-more-than-ever.md

Why second: Horthy explains why the agent goes wrong, Pocock explains why the codebase does. He ran the workflow Horthy warns about, regenerating from a spec without reading the output, and watched the system get worse on every pass. His fix is old ideas, shared design language and small verified steps, applied to an agent.

### 3. Don't Build Agents, Build Skills Instead

Barry Zhang & Mahesh Murag, Anthropic | 16:22 | AI Engineer CODE 2025
Video: https://www.youtube.com/watch?v=CEvIs9y1uog
Summary: https://aietalks.com/talks/dont-build-agents-build-skills-instead.md

Why here: the first two talks both end at the same place, write the house rules in a file the agent loads when it needs them. Zhang and Murag built that mechanism. A skill is a folder the agent only reads the name of until the moment it is relevant, which is Horthy's context problem solved at the level of the tool rather than the workflow.

### 4. From AI-Assisted to AI-Native: Building a Frontier Development Team

Clare Liguori, AWS | 20:57 | AI Engineer World's Fair 2026
Video: https://www.youtube.com/watch?v=pqlWNihgdjI
Summary: https://aietalks.com/talks/from-ai-assisted-to-ai-native-building-a-frontier-development-team.md

Why here: the first talk about a team rather than an engineer. Amazon gave 50 ordinary teams the same tool. Half got under a 3x speedup and half got a median of 4.5x, and the difference was not the tool, it was whether the team changed how it worked. Liguori's five habits are the team-scale version of what Horthy and Pocock do alone.

### 5. How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage)

Eyal Blum, Figma | 17:42 | AI Engineer World's Fair 2026
Video: https://www.youtube.com/watch?v=5Bn0xro2ol8
Summary: https://aietalks.com/talks/how-to-get-your-org-to-adopt-coding-agents-without-shipping-garbage.md

Why here: Liguori says the teams that changed pulled ahead. Blum is the one who had to make them change. His finding at Figma is uncomfortable and rings true: the strongest engineers adopt last, because they hold the undocumented knowledge and see every failure first. His answer is to fund verification before persuasion.

### 6. Your Code Has Bugs. Lean4 Has Proofs: Formal Verification for Engineers

Varun Pant, AWS | 10:07 | AI Engineer World's Fair 2026
Video: https://www.youtube.com/watch?v=lRa9sPaMyy4
Summary: https://aietalks.com/talks/your-code-has-bugs-lean4-has-proofs-formal-verification-for-engineers.md

Why last: every talk before this one ends at "a human has to check it", and every one of them admits humans cannot read pull requests as fast as agents write them. Pant's answer is to stop reading the code and check a proof instead. It only applies to the small part of a system where correctness is worth that much work, which is exactly why it belongs at the end rather than the start.

## Editor's note

Horthy says read the plan. Blum says fund verification before you fund adoption. Both assume you can see what the agent actually did, which is easy in Claude Code and hard when the agent is one you built and shipped to other people. Kitaru is our tool for that: it records agent runs so you can replay one and find the step where it went wrong. It is also why this site exists.

Written by the AIE Talks editors (the Kitaru team), not by any of the speakers.