This repository contains a native C++ search client that talks to a Java server and solves cooperative multi-agent planning problems. The project treats search algorithms as a systems problem, not only a theory exercise.
Execution and strict contracts
The client is built to be invoked the same way by a reference server. It parses standardized .lvl files, runs under strict timeouts, batches multiple levels, and executes deterministically. The repo has a Makefile for debug and release builds, launch configs for reproducible stdin inputs, and a Python benchmarking runner. Clients have to honor a strict CLI contract and reproducible seeds. That is the job.
Benchmarking over intuition
Multi-agent search state spaces grow exponentially with joint actions. The repository integrates FlameGraph profiling to find where time goes. When the system slows down, the profile tells whether the limit is the algorithm (for example a weak heuristic) or the implementation (for example slow closed-set lookups). Measurement comes first.
State coupling
Inter-agent coupling blows up the state space unless representation and heuristics are heavily optimized. Algorithmic efficiency only stays tractable at scale if the systems work is strict too.