The agent harness is solved. Synthesis is not.

Prime Intellect shipped Prime Agent, MIT-licensed, above the human-expert baseline on ARC-AGI-3. The RLM runtime is about 1,500 lines. Decomposition is a public result.

I read their source the same day. No corpus curation — web search returns five raw Google results. The word synthesis appears once, as a suggested user prompt. Sub-agent calls return a session pointer with no aggregation, conflict detection, or coverage accounting.

Recomposition is the remaining problem, and nobody is measuring it. That is the bet: the scarce work is producing what is true across a body of material that no single piece of it states, with coverage counted rather than asserted.