A single daily edition of announcements from the AI coding tools you use, read from a curated source list, summarised, and linked back to the original. No infinite feed, no account, no advertising.
Telling your agent to use TDD made it write worse code than saying nothing
Dan Luu ran roughly 80 attempts each across 26 prompt conditions, at medium and xhigh effort, asking Codex with GPT-5.6 Sol to implement the Zstd compression algorithm in Rust and scoring the fraction of runs that passed 100% of a hidden test suite. The condition with no special instructions at all beat most of the specialised ones, and the formal-methods conditions built on Verus, Lean 4, Alloy, ACL2 and TLA+ largely underperformed too. The techniques were being used in name only: agents reaching for formal methods proved properties irrelevant to the code, and the fuzzing conditions generated inputs that were overwhelmingly invalid and never explored meaningful behaviour. Large skills such as Hegel and ECC degraded results despite reading as helpful, while property-based testing through Proptest showed modest gains. Luu's own best results came from light instructions followed by adjustment after watching what the agent actually did.
Your Copilot CLI can pick a different model for each step of one task
The entry Copilot added to its model menu on 4 September is not another model: Project HydraFusion is a planner that builds an execution plan per task and spends several models on it. It picks one of three patterns — Single, where one model does the work; Cascade, where a cheaper model drafts and a quality gate decides whether to escalate; and Critique, where an independent reviewer assesses a draft and the drafting model revises once. Against Claude Opus 5 the research preview scored 4.9 points higher on TerminalBench 2.1 at 67% lower estimated cost, 1.5 points lower on DeepSWE at 36% lower cost, and 0.1 points lower on CheckpointBench at 65% lower cost. It is open to every Copilot plan — run /update, turn on /experimental, then select it — and you are billed at each model's standard token rate. GitHub says it suits first-turn, single-prompt tasks in autopilot mode; multi-turn work is not yet tuned, and intermediate drafts stay hidden until the final result.
