Concierge

Concierge's bet is that tool ordering should be a server concern, not a prompt concern. That is a defensible position: the homepage example shows a browse/cart/checkout flow where complete_purchase simply cannot be called from the browse stage because app.stages = {"browse": ["search_products"]} never exposes it. No amount of clever prompt writing beats a tool that is not on the list. The primitives are well chosen. Transactional state via app.get_state() and app.set_state() gives you atomic, consistent, session-isolated memory without bolting on a separate store. Stages are plain dictionaries, so the learning curve is a config file, not a framework. Transitions being self-discoverable means the agent is told what it can do next rather than guessing, which is exactly where emergent tool-calling goes wrong. Semantic tool search is the release valve for large surfaces. Above roughly 100 tools, staged workflows stop being enough, and enabling provider_type=ProviderType.SEARCH with max_results collapses everything into search_tools and call_tool. That is a pragmatic escape hatch rather than a claim that staging scales indefinitely. Where we would push back: the headline numbers (78% token reduction, 65% fewer errors) are the vendor's own and untested by us. Support is Python-only, so a polyglot team gets reliability on one service and none on its TypeScript neighbours. And because the layer sits on an open protocol standard, protocol-level changes can absorb or duplicate parts of what it does. None of that is disqualifying — it just means Concierge is a middleware bet you should size against your actual stack.