Late
A CLI coding agent for long-horizon tasks that isolates execution noise from the model's context, aimed especially at running well with local models in a small context window.
Late is a single, statically compiled binary with no Python or Node.js dependency that works with any cloud provider or local model, automatically finding an already-running llama-server. It targets a specific architectural problem: standard coding agents let codebase scans, compiler errors, file reads and retries accumulate directly in the primary model's context, which the project says degrades reasoning quality once context use crosses 40 to 50 percent, citing published research on long-context reasoning collapse and on wasted self-reflection tokens.
Late's approach splits the work and treats the central context as a scarce resource, isolating execution steps so the model's working context stays clean during long, multi-step workflows such as resolving a complex merge conflict. It reports handling more than 200,000 tokens of work inside a 64,000-token context window, and can run a full local development setup in around 5 GB of video memory.