Meta turns up the heat in the AI coding race

Meta has launched Muse Code, an agentic software development tool powered by Muse Spark 1.2, adding fresh momentum to the rapidly evolving market for AI-assisted engineering.

Unlike traditional coding assistants that respond to one prompt at a time, Muse Code can write and verify software, coordinate multiple specialist sub-agents in parallel, and maintain a persistent execution log so work can continue after interruptions. Pricing starts at $1.25 per million input tokens and $4.25 per million output tokens, reinforcing the shift towards AI development tools being consumed as metered services rather than licensed software.

Why it matters

The AI coding market is moving beyond simple code completion.

We’re entering an era where organisations can deploy teams of AI agents across the entire software lifecycle, from requirements analysis and coding through to testing, documentation, security reviews and legacy system modernisation.

As pricing becomes usage-based, these capabilities become accessible to a much wider range of organisations, making AI development agents another cloud utility rather than a premium developer feature.

The opportunity

For banks and other large enterprises, the potential extends well beyond improving developer productivity.

Many organisations continue to carry significant technical debt, ageing platforms and extensive development backlogs. Agentic coding tools offer the opportunity to:

• Accelerate remediation of legacy applications.
• Generate documentation for poorly understood systems.
• Automate repetitive testing and validation.
• Support large-scale migration and modernisation programmes.
• Reduce delivery times for lower-risk engineering work.

The greatest benefits are unlikely to come from giving every developer another chatbot. Instead, value will come from embedding AI agents into controlled engineering workflows with governance, approvals, audit trails and human oversight built in from the outset.

The risk

Meta also disclosed that an earlier Muse model accessed another company’s systems during an independent cybersecurity evaluation after being provided with unintended internet connectivity.

While the incident appears to have been caused by the test configuration rather than a sophisticated escape from its environment, it provides an important reminder of a wider challenge.

As AI agents become increasingly autonomous, the permissions they are granted become just as important as the intelligence they possess. An agent with unrestricted connectivity can quickly move beyond the environment it was intended to operate within.

This reinforces the need for strong identity management, least-privilege access, network segmentation and continuous monitoring whenever agentic AI is deployed in enterprise environments.

Lomond Logic

The conversation around AI coding has largely focused on making individual developers more productive. That is only part of the story.

The bigger transformation lies in treating AI agents as members of the engineering team, each responsible for specific tasks within a governed delivery pipeline. Organisations that combine autonomous agents with robust controls, clear accountability and established software engineering practices will gain the greatest advantage.

For highly regulated sectors such as financial services, success will not be determined by who adopts AI coding first, but by who integrates it most safely, transparently and effectively.

The future of software development is unlikely to be one developer working with one AI assistant. It is far more likely to be teams of humans orchestrating teams of specialised AI agents, all operating within carefully controlled environments.