Sig AdvisorySig Advisory
agents

VS Code 1.143 lets you run one coding task on several models at once and pick the winner

VS Code 1.143 makes the Copilot harness the default for new agent sessions and adds a preview Compare Models mode in the Agents window, plus experimental on-device models and worktree cleanup.

Written by Mara at Sig · Oct 22, 2026

Compare Models runs the same prompt as separate attempts, each in its own session and worktree. You can mix Copilot, Claude and Codex harnesses and set different reasoning levels. You can optionally pick a Judge model that scores the attempts, with time and token use for each, and then hit Synthesize to build on the winner with ideas from the others. Turn it on with sessions.chat.compareAgents.enabled in the Agents window.

The on-device option runs Microsoft's MAI Code 1.1 Flash locally through llama.cpp, and those responses don't use AI credits. The catch: it needs an NVIDIA GPU with 96 GB or more of GPU memory, so most laptops are out.

A new worktree cleanup screen (Chat: Open Worktree Cleanup) shows how much disk old agent session worktrees use and lets you delete them. Microsoft's own table shows a long chat with about 530 MB of history going from 35.6 seconds to 9 seconds to open cold.

Why it matters

Running two or three models on the same real task is the fastest way to find out which one fits your codebase, and until now you had to set that up by hand. Each attempt costs its own tokens, so the judge's token count per attempt is worth reading.

Try this

In VS Code 1.143, enable sessions.chat.compareAgents.enabled, open the Agents window, choose Compare Models in the model picker, select two models, and give them one real bug. Check the judge's scorecard for time and tokens per attempt.

Sources

Back to all newsGet Sig Weekly every Monday