Extreme LLM Engineering: Claude Opus 5.5 Fuels Multi-Week, High-Compute Agent Workflows for 'Impossible' Projects

A prominent developer has unveiled an unorthodox and resource-intensive approach to software development, pushing Anthropic’s Claude Opus 5.5 to its perceived limits for projects previously considered ‘impossible.’ Citing Claude’s capabilities as ‘addicting,’ the developer detailed orchestrating agent loops that operate continuously for weeks, consuming thousands of hours of LLM inference and incurring compute costs that can dwarf token expenditures. This extreme methodology, termed ‘hill climbing’ with verifiable rewards (RLVR), has led to remarkable outputs such as ‘TS Rust,’ a TypeScript compiler rewritten in Rust achieving 4-10x speed improvements, and ‘Lust,’ a Rust compiler replacement aiming for 100x faster compilation for specific subsets. Perhaps most strikingly, Claude autonomously developed ‘Skate 3 Bench,’ a playable 3JS-based game engine replicating EA’s Skate, a project that reportedly consumed nearly $5,000 in inference over 12 billion tokens with minimal human intervention.

The core strategy centers on decomposing complex tasks into highly tedious, programmatically verifiable sub-goals, enabling agents to iteratively refine their output. Verification mechanisms range from precisely matching compiler outputs to utilizing LLM-judged screenshot comparisons for graphical fidelity in game development. Crucially, the process highlighted significant bottlenecks in compute provisioning, with agents spending hours awaiting resources. Solutions involved scaling local compute resources, integrating CI platforms like Blacksmith, and most notably, leveraging Boat.dev for on-demand ephemeral VMs. This dynamic compute allocation dramatically reduced ‘gate times’—the period agents spend waiting for build and test cycles—from upwards of 70 minutes to just 6 minutes, exponentially accelerating project progress. The findings suggest a new developer paradigm: a ‘manager’ of autonomous agents, focused on identifying and mitigating environmental blockers to unlock unprecedented productivity, albeit at substantial financial and cognitive investment.