The cost-tiering question for AI pipelines usually lands as: which agents can run on a cheaper model? The honest audit answer is: none of them.
i.Why agent-level tiering fails
The appeal is straightforward. You have a roster of specialized agents. Some of them, the security reviewer, the research scout, the editor, seem more demanding than others. Route the demanding ones to a capable model. Route the rest somewhere cheaper.
The problem is that “the rest” is not what you think it is.
A data-cleaning agent does judgment work when it encounters an ambiguous field. A formatting agent does judgment work when the source material does not match the expected structure. A summarizer does judgment work when it decides which details are load-bearing. Agents are not mechanical at the task level. They are mechanical or not at the dispatch level, which is a finer distinction.
An agent that is mostly mechanical will encounter judgment moments. Downtiering the whole agent to save on the mechanical fraction means the judgment moments run cheap. That is where quality degrades, and it degrades silently.
ii.The granularity that works
The unit of tiering is the dispatch, not the agent.
An agent may have a research phase, a synthesis phase, and a formatting phase. The research phase is retrieval with some filtering. The synthesis phase is judgment. The formatting phase is mechanical. Routing each phase to the appropriate model tier is the decision that actually optimizes cost without trading quality.
In practice, this means two things.
First, document what each dispatch is actually doing before assigning it a model. “This agent summarizes” is not enough. “This dispatch extracts the three most consequential decisions from a structured handoff file, where the file has a predictable schema and the extraction criteria are explicit” is tierable. “This dispatch synthesizes research from open-ended web retrieval and weighs competing claims” is not.
Second, build deterministic gates at hand-back. If a cheaper model handles a dispatch, the gate at the output boundary catches quality failures before they propagate. The protection is structural, not assumed. A cheap model that produces correct output on easy cases and flawed output on hard cases is manageable if the hand-back gate sees the hard-case output and flags it. It is not manageable if the gate assumes the model was capable enough.
iii.The CAIO decision
The rule that emerged from this audit: tier the individual dispatch whenever a purely mechanical sub-task can be isolated. Quality is protected by deterministic gates at hand-back, not by model strength alone.
The implication for teams building agentic pipelines: cost audits should map dispatch types, not agent names. An agent might run fifty dispatches across a project. Some of those dispatches are mechanical. Most are not. Knowing which is which tells you where the savings actually live.
The alternative, the agent-level reassignment, optimizes the wrong unit and accepts silent quality loss in exchange for broad cost reduction. That tradeoff is usually not visible until a downstream artifact surfaces a degraded output, at which point the cost “saved” has been spent on remediation.
Drafted with Bishop, my AI partner.
Words picked, edited, and approved by me.
Model provenance: Claude Code (Claude Opus and Sonnet)