✶Explainer
The moment coding agents stopped being assistants
Naval dates the shift to around December 2025 and the release of Claude Opus 4.5, when agents began staying on track across whole builds rather than returning snippets. The old model was a coding assist: ask for a fix, receive a pile of code, paste it into your IDE. The new model is a long-lived agent that builds an app end to end. He had tried earlier agents with mixed results, but says this time the hype felt real.
- December 2025 and Claude Opus 4.5 marked the inflection point
- Agents stay on track and build apps soup to nuts rather than emitting snippets
- The experience resembles having a fast, free junior programmer
- Naval has a computer science degree but had not seriously coded in decades
- Activation energy, not ability, was the old barrier to writing code
“So around December of 2025, the coding agents in AI hit an inflection point with the release of Claude Opus 4.5.”
“And the activation energy to writing code is really high.”
#ai-agents#vibe-coding#developer-tools
✶Explainer
Why coding agents are really Unix programs wearing English
Naval explains the architecture behind the terminal-first workflow. Agents are trained on text tokens, and most of the code they trained on was Unix, so they are natively fluent in the shell. They run in a command line interface, execute commands, reach the file system, chain grep, awk, sed and pipes, run cron jobs to stay long-lived, and spawn additional shells and tasks. Modern operating systems are Unix underneath anyway, with macOS famously built on BSD.
- Agents are text in, text out, which is exactly what Unix is
- Most training code on GitHub and Stack Overflow was Unix
- They connect to the shell, the file system and standard commands
- Cron jobs make them long-lived; spawned shells make them parallel
- The AI acts as a translator from English into the tool jargon
“So these agents are just long-lived coding AIs that are connected to Unix at a core level.”
“So underneath these are all Unix, which is all text in, text out.”
#unix#cli#ai-agents#architecture
✶Explainer
Every recent AI leap is an error-correction trick
Naval reframes the last few years of AI progress as successive applications of error correction. Agents error correct and learn, some of them watching YouTube videos at night to improve at tasks assigned during the day. Reasoning models turn next-token prediction into a pseudo-thinking process that error corrects at each step. Reducing hallucinations was error correction too. He wonders aloud what the next application will be, and floats agents correcting each other, while doubting the analogy holds.
- Agents go out, error correct, and improve their own skills
- Thinking models are next-token prediction with step-wise error correction
- Hallucination reduction was itself an error-correction process
- The open question is what error correction gets applied to next
- Agent-to-agent correction is the obvious candidate and the weakest analogy
“The most interesting thing to me about agents is their ability to error correct and learn.”
“Getting rid of hallucinations was also an error correction process.”
#ai-research#reasoning#agents
✶Explainer
Why a growing codebase breaks a coding agent
Naval explains the mechanical reason agents degrade on large projects. State of the art is roughly a million tokens of context, which he says will look laughable in future, and the transformer attention mechanism scales as the square of the token count. Once the codebase exceeds what fits, the model starts guessing, approximating and compacting. The symptoms are recognisable: losing the plot, fixing the wrong thing, and fixing the same bug five times.
- About a million tokens is the current ceiling, roughly a million words
- Attention cost scales as the square of context length
- Past the ceiling the model compacts, guesses and approximates
- Repeat fixes on one bug are the diagnostic signal
- Guidance shifts onto the operator as complexity grows
“And right now, the state of the art is about a million tokens, which will be considered laughable in the future.”
“They start losing the plot, they get lost. They start fixing the wrong thing. They fix the same bug five times.”
#context-window#transformers#ai-limits#codebase
✶Explainer
Great engineers' taste is now inside the models
Naval offers two explanations for the recent jump in coding model quality. One is near-recursive training, where one model helps improve the next. The bigger one, he suspects, is that many of the best software engineers started using these tools in the last few months, so the models absorbed not just their code but their judgement about what is good. He generalises: improving models requires high-taste feedback loops, which are far harder to build than they look.
- Recursive training, where one model improves the next, is part of it
- The larger factor may be top engineers adopting the tools en masse
- The models gained their code and their taste at the same time
- High-taste feedback loops are the real scarce input
- Some domains make such loops tractable; others do not
“But I think the bigger reason might just be that a lot of the best software engineers started using these models in the last few…”
#model-training#taste#feedback-loops