Design the seam: A simple rule for what your AI agents should (and shouldn't) decide
The most important decision in any agent rollout isn't which model to use. It's where humans step in. A three-tier framework you can apply this quarter.
Most agent pilots start with the wrong question: Which model should we use?
Models matter. But by Andreessen Horowitz's measure, the cost of a fixed level of model quality fell roughly 1,000x between November 2021 and November 2024, and the gap between the frontier and the next tier has narrowed every quarter for two years. The model is becoming a commodity input.
The question that decides whether your pilot ships or blows up is different: Where does the agent stop and a human take over?
We call that boundary the seam. You may know it as human in the loop. In most deployments, nobody designs it. It just emerges, and the widely reported outages of the past year all broke there.
The three-tier rule
The rule is simple: the easier a decision is to undo and the lower the stakes, the more your AI agents should decide on their own. For every pod (a small team of humans and agents working toward one outcome), sort each decision into one of three tiers before the agents go live:
- Tier 1: reversible and low stakes. The agent decides and acts on its own. For example: drafting outreach, routing a ticket, summarizing a call.
- Tier 2: reversible but high stakes. The agent proposes, and a person approves before it runs. For example: sending a large discount, changing a customer record, publishing externally.
- Tier 3: irreversible. A named person decides, however confident the agent is. For example: deleting production data, moving money, signing a contract.
Two details make this work. First, the boundary is enforced in code, before the action runs, not in a policy PDF reviewed after the fact. Second, Tier 3 belongs to a named person, not a committee.
Put the seam in every layer of the stack
Production agents run on five layers: intelligence, action, governance, coordination, and economics. Each one has its own seam:
- Intelligence: people decide what goes into the retrieval index and which evaluations a model must pass.
- Action: people set the approval boundary, the three tiers above.
- Governance: one person writes the policy, and a different person audits that the runtime enforces it.
- Coordination: people bring judgment, taste, trust, and ethics. Agents bring speed, parallelism, and 24/7 availability.
- Economics: one person owns price per completed task as a managed metric.
Measure the right things
An agent can generate a hundred actions a minute. Reviewing activity at that pace defeats the point. Measure outcomes instead:
- Task completion rate
- Intervention rate (how often a human has to step in)
- Price per completed task, not cost per token
- Time to recovery when something breaks
If the intervention rate falls and stays low, the seam can move toward more autonomy. If you deploy into higher-stakes work, it should move back.
Start with one pod
You don't need to redesign the company to begin. Pick one imperative, assemble one pod, and ship one outcome. Good first pods include outbound sales, tier-one support, sales development, recruiting, and financial planning. Keep it small enough to measure in weeks and bounded enough that a failure won't hurt the business.
Then name a seam owner: one executive accountable for what agents may do without asking, what they must ask about, and what they must escalate. Early on, that's often the CEO. In a larger company, it may be the VP of the business unit where the pod runs.
Get the full playbook
The white paper breaks down all five layers, the four SuperOrg principles, and a three-phase roadmap for going from a first pod to a new operating model in 36 months.