dotey on Astra vs Codex: models can internalize how to use tools, never the harness itself
dotey · x · 2026-09-06
Responding to claims that 'Codex can't handle Astra anymore, the coding harness has been internalized,' dotey lays out a detailed counterargument:
- Models internalize strategy — when to call tools, how to decompose tasks, how to recover from failures — but the model only outputs tokens; actual execution, context management, permissions, sandboxing, and recovery all live in the harness, which cannot be trained into weights
- Analogy: a great driver without a car can only mime steering
- As models improve, the harness gets thinner on 'how to think' but faces higher demands on execution, permission control, and state management
- Conclusion: the model drives the computer through the harness; if Astra outgrows Codex's design, Codex should be upgraded — not dropped. Model and harness are complementary.
More from coding & agent
- Scaling agent code means humans become PCP checkers, argues developer — willcb · 2026-09-06
- Dev argues massively scaling agent-written code is feasible if it stays human-legible — willcb · 2026-09-06
- Running 95 sandboxed agent containers: lessons from the first paying user — aihudagency · 2026-09-06
- Being a senior dev in 2026 means explaining to AI why we don't do that — josh_wills · 2026-09-06
- rauchg goes big on WebMCP: agents riding existing web infrastructure — jeff_weinstein · 2026-09-06
- User rebuilds VR hit Gorilla Tag with GPT-6 Astra, claims better graphics — Acid_God_ · 2026-09-06