Meta's AI Ambitions Grow with New Terminal Coding Agent

Meta, considered a bit of a straggler in the AI harnesses realm, is making strides to catch up. This week, the company released a new terminal coding agent aimed at programmers looking for assistance with complex tasks across large software code bases.

Meta's AI Chief Touts Cost-Effectiveness of Muse Code

Muse Code, which is currently available in beta, can accomplish "complete software engineering tasks across large repos," Meta CEO Mark Zuckerberg said in a social media post on Wednesday. Those tasks include "planning changes, writing code, validating the results," he added.

The move attempts to position Meta more competitively, and more affordably, with AI Lab peers like OpenAI and its coding agent Codex, and Anthropic with Claude Code.

Muse Code is powered by Meta's previously released coding model, Muse Spark. It handles large projects by launching its own agents, which then work simultaneously.

• When a job is big enough, it fans out to separate sub-agents working in parallel in isolated worktrees.
• Your working copy is never touched.

In testing, Meta's AI team had Muse Code build six features for a game simultaneously with no collisions.

"We think that for a lot of workflows and a lot of use cases, this can be an incredibly good option, especially from a cost perspective," Alexandr Wang, Meta’s AI chief, told the Wall Street Journal.

Meta has been attempting to grow its AI presence by pouring money into development. In June, it expanded beyond its core focus of using AI to support its advertising business and entered the enterprise AI market with an agent aimed at customer service and support.