mas-compiler

Use the three-phase compilation: decompose each agent's role into atomic capabilities, assign each capability an external tool or internal policy backend, then internalize the communication topology by embedding handover output constraints into each phase's policy.

Architectures with cyclic communication like debate loops violate serializability, agents with private state or hidden objectives violate shared history, and mixed-model systems violate the homogeneous backbone condition. Adversarial negotiation and true parallel sampling are also non-compilable.

Compiled single agents maintain accuracy within about ±4% of the original MAS according to the referenced paper's benchmarks. They reduce token usage by roughly 54% and latency by about 50% on average.

Compilation produces a directory containing a SKILL.md with one [PHASE] section per capability in topological order, plus a tools/ directory with one Python file per externalized capability. Each tool returns a flat dict and wraps errors in try/except.

Compilation is not beneficial when skill selection cost exceeds the communication overhead eliminated, which occurs when skill libraries grow beyond roughly 50-100 entries. Selection accuracy collapses near 200 skills due to a scaling law phase transition.