Fetching from the wire…
01
02
03
04
05
06
07
08
Top 5 · 2026-03-31 · source-backed
I've been saying for months that the real gains aren't in switching models. They're in how you set up the environment around the model. Now there's quantitative proof.
Stanford IRIS Lab published Meta-Harness, a system that autonomously evolves its own coding harness, system prompts, tool definitions, completion-checking logic, all of it, by reading per-task execution traces. On TerminalBench 2.0 (89 Dockerized tasks, 5 trials each), Meta-Harness running Claude Opus 4.6 hit 76.4%, beating the hand-engineered Terminus-KIRA baseline at 74.7%. On Haiku 4.5, it scored 37.6% vs. Goose at 35.5%.
The key insight is environment bootstrapping. Before the agent loop starts, Meta-Harness snapshots the sandbox: working directory, available languages, package managers, memory constraints. It injects all of this into the initial prompt. This eliminates 2-5 early exploration turns that agents normally waste on ls, which python3, cat package.json. That's not a minor optimization. That's the difference between an agent that understands its environment and one that's groping around blind for the first five minutes.
The proposer is itself a Claude Code agent that uses grep and cat to diagnose failure modes across up to 10 million tokens of diagnostic context per optimization step. The harness writes its own improvements. Read that again.
This connects directly to something Georgi Gerganov said while reflecting on llama.cpp hitting 100K GitHub stars: the main issues users face with local models "mostly revolve around the harness and some intricacies around model chat templates and prompt construction," not model quality. The pipeline from input to output involves components "developed by different parties" that are "not only fragile" but lack cohesion.
And then there's this from Latent Space: "Opus scores ~20% higher in Cursor than Claude Code." Same model. Different harness. 20% performance gap.
If you're spending time evaluating which model to switch to, stop. Spend that time on your system prompt, your tool definitions, your environment setup. That's where the gains are. The Meta-Harness paper proves it with numbers.
Each link below shares sources, entities, or timing with this story.
Claude Code uses Opus / Shared entities / Same source domain / Shared topic / What happened next
Linked by a graph relationship (Claude Code uses Opus); both cover Claude Code, Cursor, Different, Same; reported by the same outlet (github.com).
Linked by a graph relationship (Claude Code uses Opus); both cover Claude Code, GitHub, Meta, Opus; reported by the same outlet (github.com).
Opus built by Anthropic / Shared entities / Same source domain / Shared topic / What happened next / Tension
Linked by a graph relationship (Opus built by Anthropic); both cover Claude Code, Claude Opus, GitHub, Opus; reported by the same outlet (simonwillison.net).
Claude Code uses Opus / Shared entities / Same source domain / Shared topic / What happened next
Linked by a graph relationship (Claude Code uses Opus); both cover Claude Code, Claude Opus, Opus, Same; reported by the same outlet (latent.space).
Opus built by Anthropic / Shared entities / Same source domain / Shared topic / What happened next
Linked by a graph relationship (Opus built by Anthropic); both cover Claude Code, Cursor, Latent Space, Spend; reported by the same outlet (latent.space).
Claude Code uses Opus / Shared entities / Same source domain / Shared topic / What happened next
Linked by a graph relationship (Claude Code uses Opus); both cover Claude Code, Cursor, GitHub, Same; reported by the same outlet (github.com).
Stanford benchmarked against DeepSeek / Shared entities / Same source domain / Shared topic / What happened next / Tension
Linked by a graph relationship (Stanford benchmarked against DeepSeek); both cover Claude Code, Different, Same; reported by the same outlet (github.com).
Opus built by Anthropic / Shared entities / Same source domain / Shared topic / What happened next
Linked by a graph relationship (Opus built by Anthropic); both cover Claude Code, Latent Space, Same; reported by the same outlet (latent.space, simonwillison.net).