Are We Over-provisioning AI Agents by Default?

I've been thinking about model selection in multi-agent systems, especially as frontier models become increasingly optimized for coding and long-running software-engineering work. Those models are extraordinary. But a lot of agent work isn't sophisticated coding—or even deep reasoning. Think about the calls inside a typical workflow: classify or route a message; extract a few structured fields; summarize a document; translate or rewrite text; search documentation or logs; compare records against predefined criteria; call a well-defined API; validate output; update a database or CRM; handle retries, bookkeeping, or status tracking. A planner or reviewer may genuinely need frontier-level reasoning. A classifier or extractor often doesn't. That made me wonder whether we're thinking about model selection the wrong way. Instead of asking: “What's the smartest model I can use?” maybe the engineering question should be: “What's the least expensive model that reliably clears the quality threshold for this particular task?” A rough cost example Suppose a multi-agent system makes 10,000 model calls per day. For a deliberately simple comparison, assume each call averages: 2,000 input tokens; 500 output tokens. Using current standard short-context pricing for three models in the same family, that gives roughly: flagship: $0.025 per average call; main: $0.010; economy: $0.001. Now compare three architectures. 1. Everything goes to the flagship model 10,000 calls/day → about $7,500/month. 2. 80% main / 20% flagship Same number of calls → about $3,900/month. 3. 60% economy / 30% main / 10% flagship Same number of calls → about $1,830/month. That's a 75.6% difference from the all-flagship setup under these particular assumptions. Obviously that isn't a prediction or a universal routing recipe. Different workloads have very different token distributions and quality requirements, and real costs also depend on caching, retries, context size, reasoning tokens, tool calls, etc. More importantly, the idea isn't to accept worse results just to save money. The assumption is that cheaper models only get tasks where they already meet the required quality and reliability threshold, while difficult, ambiguous, high-value, or high-risk work gets escalated. The coding angle seems important too Frontier models are increasingly being pushed hard on coding, agentic coding, terminal work, tool use, and long-horizon software engineering. That makes sense economically and technically. But if an agent system spends a large share of its calls summarizing, classifying, extracting, translating, routing, monitoring, calling deterministic tools, or doing other structured operational work, then benchmark leadership on difficult coding tasks doesn't necessarily translate into proportional value on every one of those calls. So I'm increasingly thinking of model capability as an infrastructure resource—something to allocate according to task difficulty, uncertainty, value, and risk, rather than a single default chosen for the whole system. For people running real agent workflows: Do you route different tasks to different model tiers, or still use one strong model for almost everything? And if you do route them, what signals have actually worked well for deciding when a task should be escalated to the frontier model?