Evolution Arena — Simulation Engine

The problem, in business terms

Build an evolutionary simulation engine where thousands of agents compete, eat, breed, and die — each one executing a program written in a custom domain-specific language. That means building a real language toolchain (lexer, parser, AST, interpreter, validator) plus the world simulation around it.

The statement of work

The statement of work specified the DSL, the world rules, deterministic state-hash verification, and per-capability test coverage — the hardest tier of spec in our benchmark suite.

What the agent team did

• —A two-agent team built the complete DSL toolchain: lexer, parser, AST, interpreter, validator, and unparser
• —The world subsystem — genomes, agents, field, lifecycle — was built on top with deterministic, seedable behavior
• —Correctness is verified by deterministic state-hash checks: the same seed always produces the same world
• —Every capability area got its own test suite, per the spec

What shipped

• —36 Python source modules across the language toolchain and the simulation engine
• —883 automated tests across 50 test files
• —Deterministic simulation verified by state hashing — reproducible to the byte
• —Full documentation of the DSL and the engine architecture in the repository

Verify it yourself

• Open the code: the public repository on GitHub
• Watch it being built: replay the actual run
• Check the independent score: this run on the public benchmark