Building a multi-agent content engine

Coordination failure between agents is the real system killer, ahead of hallucinations and token burn.

Instrument the handoffs and log the constraints passed between agents, not just raw output.

Put guardrails and fallback logic at the router, not inside individual agents.

Forecast failures early, ProMAS style, and use a surrogate eval loop, AgenticGEO style, instead of end-to-end tests for every tweak.

My agentic orchestration approach draws on two papers, ProMAS and AgenticGEO.

ProMAS paper (arXiv 2603.20260)

AgenticGEO paper (arXiv 2603.20213)