Megaplan harness edges Opus on SWE-bench

This is a harness story more than a model story: the claim is that structured planning, critique, gating, and review can unlock much better coding performance from open models than one-shot execution.

• –The live setup uses GLM-5.1 for prep, plan, execute, and review, with MiniMax-M2.7-highspeed handling critique and review, which is a concrete example of phase-specialized orchestration
• –The repo frames Megaplan as a reusable workflow layer, not a one-off benchmark script, which makes the result more interesting for agent builders than for raw model rankings
• –The result is still early and noisy: 26 scored tasks is a small slice of SWE-bench Verified, so the lead could move as the remaining 474 tasks come in
• –The fact that all code and data are public makes this unusually replicable for a leaderboard claim, which should help separate signal from hype
• –If the curve holds, this strengthens the case that better agent scaffolding can matter as much as marginal model gains on software engineering tasks