evm-bench benchmarks are (typically) expensive Solidity contracts paired with configuration.
Benchmarks are built independently of any runner using solc running in Docker. The evm-bench framework picks up on benchmarks by scanning for benchmark.evm-bench.json files, which have a schema. That schema has more information on the structure of benchmark metadata file.
You want to first start off by creating a new Solidity contract. This can be whatever you want it to be, but presumably it will be expensive to run in an EVM. Then figure out the calldata you need to execute your benchmark. Do not have your benchmark be in the constructor. The constructor is not benchmarked by runners. The runners will benchmark the time it takes to call the contract with the calldata you supply.
All you need now is a new benchmark.evm-bench.json file somewhere under this directory (since this is where the tool scans for benchmarks by default). Use the other benchmarks here as an example! Create a new folder and add resources under that folder. Note that if you plan to share resources among benchmarks (e.g. a shared Solidity library), make sure the benchmark metadata has the correct build context. See benchmarks under erc20 for an example of this.
Once you have your benchmark, it's time to test! Consider running the evm-bench framework with a single runner (revm is the most stable in my experience) against your new benchmark to start, then move on to running it on all runners. It would look something like RUST_LOG=info cargo run -- --runners revm --benchmarks <my_new_benchmark_name>, if you need more information about logs you can tweak RUST_LOG.