Lovable Extends Autonomous Runtime to 10-Hour Work Sessions.

Lovable can now continue working autonomously for up to 10 hours, giving its coding agent enough time to tackle tasks that involve research, experimentation and repeated technical validation. The company demonstrated the longer runtime by asking the agent to build a neural network that identifies strummed guitar chords, with no training data supplied by the user.

According to the LinkedIn announcement, the entire workflow ran inside Lovable’s sandbox. The agent generated data, located a public dataset, assembled a training pipeline and iterated on the model before producing a browser-based application.

That is a broader workload than the interface-focused app generation commonly associated with vibe coding tools. It also offers a useful test of whether longer-running agents can do more than simply spend additional time writing code.

Chord recognition is difficult because recordings vary by instrument, tuning, playing style, microphone and background noise. The user’s constraints added another challenge: the model needed to be fast, accurate and small enough to run in a browser.

Lovable reportedly generated 23,000 synthetic training samples and found GuitarSet, a public guitar audio dataset, without the user supplying either source. It then created holdout and validation sets, augmented the data with noise and detuning, and trained several iterations.

The final system used a transposition-tied multilayer perceptron, or MLP. Chroma rotations allowed it to cover all 12 musical roots without requiring a separate model for each key. Lovable also combined the neural network with a template-based method in a 70/30 ensemble, rather than relying entirely on one prediction approach.

The resulting neural model contained 2,185 parameters. The announcement describes it as smaller than the application’s 8.5 KB favicon and says inference completes in 38 microseconds in the browser. Those figures come from the demonstration rather than an independent benchmark, but they show the kind of optimization a long-running agent can attempt when it has time to compare multiple approaches.

A 10-hour ceiling could make Lovable more useful for work that cannot be completed in a short code-generation pass. Potential examples include large migrations, test-suite repair, dataset preparation, dependency upgrades and performance tuning. These jobs require an agent to inspect results, revise its assumptions and try again.

The guitar demo highlights several capabilities that benefit from a longer horizon:

• Independent data preparation, including synthetic sample generation and dataset discovery
• Iterative experimentation, with multiple training runs and validation steps
• End-to-end execution, from pipeline setup to browser inference
• Constraint-driven optimization, balancing model size, latency and recognition quality

Long sessions also increase the importance of visibility and cost controls. Lovable’s August 25 changelog introduced a way to inspect the elapsed time and credits used by a message while it is still running. That feature is particularly relevant when a single instruction can remain active for much of a working day.

There are still open questions. The announcement does not identify which subscription plans receive the 10-hour runtime, whether every task can run for the full period, or what safeguards govern external data and compute use. The chord recognizer is also one successful demonstration, not evidence that every underspecified machine-learning task will work on the first attempt.

For small teams, the practical shift is nevertheless significant. Lovable is positioning its chat interface not only as a way to assemble an application, but as an environment where an agent can conduct a lengthy technical workflow with limited intervention.

What is Lovable’s 10-hour autonomous runtime?

It is the ability for Lovable to continue working on a user’s request for up to 10 hours. This gives the agent more time to research, write code, run experiments, validate results and revise its approach without requiring a new prompt after each step.

What did Lovable build in the demonstration?

Lovable created a browser application that recognizes strummed guitar chords. It generated 23,000 synthetic samples, found the GuitarSet public dataset, trained several model iterations and produced a 2,185-parameter model that reportedly runs in 38 microseconds.

Which Lovable plans include the longer runtime?

The announcement does not specify eligible plans, additional pricing or usage limits. Lovable’s changelog separately confirms that users can inspect the time and credits consumed while a long-running message is in progress.