Should This Be an Agent?

Overview

The spectrum gave you positions. This topic is the gate you pass before moving right along it.

Four criteria. All four must hold. If any one fails, the correct answer is a simpler position — and the most common reason projects go wrong here is treating a failed criterion as an obstacle to work around rather than as an answer.

The four criteria

graph TD
  A{"<b>Complexity</b><br/>multi-step, and hard to<br/>fully specify in advance?"} -->|no| S["<b>Stay simpler</b><br/>chain, router,<br/>or single call"]
  A -->|yes| B{"<b>Value</b><br/>does the outcome justify<br/>higher cost and latency?"}
  B -->|no| S
  B -->|yes| C{"<b>Viability</b><br/>is the model actually<br/>capable at this task?"}
  C -->|no| S
  C -->|yes| D{"<b>Cost of error</b><br/>can mistakes be caught<br/>and recovered from?"}
  D -->|no| S
  D -->|yes| G["<b>Agent, bounded</b><br/>proceed to the<br/>constraints"]

  classDef gate fill:#E8A33D,stroke:#C4832A,stroke-width:2px,color:#ffffff
  classDef simple fill:#7D9A4E,stroke:#5F7A38,stroke-width:2px,color:#ffffff
  classDef go fill:#D9695B,stroke:#B4483A,stroke-width:2px,color:#ffffff
  class A,B,C,D gate
  class S simple
  class G go

Four gates in series. Every “no” exits to the same place, and that exit is a successful outcome rather than a setback.

Complexity. Is the task genuinely multi-step, with a shape that cannot be fully specified in advance? “Turn this request into a scheduled visit, handling whatever turns out to be in the way” qualifies. “Answer a question from the documentation” does not — it is two steps and you know both of them.

Value. Does the outcome justify the higher cost and latency? Agency is not free, and a task worth a few cents of value does not support a loop that might cost a dollar.

Viability. Is the model actually capable at this type of task? Module 4 made the point that neither retrieval nor training fixes incapability; neither does iteration. An agent that cannot do the task reliably in one step will not become able to do it by being allowed eleven. It will produce a longer, more expensive, more confident failure.

Cost of error. Can mistakes be caught and recovered from? This is the criterion that fails most often in the systems people most want to build.

The asymmetry nobody mentions

Here is the uncomfortable structure of the four criteria.

Agency is most attractive precisely where the work is consequential — where the system takes real actions, changes real state, and saves real money by not needing a human. But that is exactly where the fourth criterion is hardest to satisfy, because a wrong action is not a wrong sentence. It is a credit that left the building, a job moved on a technician’s calendar, a device reset while someone was using it.

So the appeal of agency and the safety of agency are inversely correlated. That is not a reason to refuse it. It is the reason the rest of this section exists: if you cannot satisfy the fourth criterion as stated, you can often engineer it into being satisfied — reversibility, confirmation gates, idempotency, audit trails — and then the answer becomes yes, bounded.

What you cannot do is note the criterion and proceed anyway.

Genuinely considering the pipeline

The rubric for this module asks that the simpler alternative be genuinely considered, which is easy to claim and easy to fake. A technique that makes it real:

Write the pipeline version first, in five lines or fewer. A fixed sequence of steps, with the branches enumerated. Then answer one question: what can this not do?

Three outcomes, all useful. If you cannot name anything it fails at, build the pipeline — you have just saved a quarter. If the only things it fails at are rare, build the pipeline and route the rare cases to a human. If it fails at something common and important, you now have a specific, defensible justification for agency, phrased as a capability gap rather than a preference.

The cost of this exercise is an hour. The cost of skipping it is discovering in month four that a fixed sequence would have done it, after you have already paid the non-determinism, latency, cost-variance and debuggability bills from the previous topic.

The simplest thing is often not what you think

One more observation, because it reframes the decision.

The question “should we build an agent?” is usually asked as though the alternatives were an agent or nothing. The real alternatives include: a chain; a router with three branches; a chain plus a human for the residue; and a chain that prepares work for a human rather than completing it. That last one is frequently the highest-value, lowest-risk design available, and it almost never gets proposed because it is unglamorous.

The gate applied to both paths

Answering a documented question. Complexity — fails. It is retrieve, generate, check: two or three known steps. That is the end of the assessment. Three criteria go unexamined because the first one settles it, and the answer is a chain.

Taking an action. Complexity — passes. Rescheduling requires establishing which job, checking entitlement, making the change and verifying it, and what is in the way varies per request. Value — passes. The actions are precisely what make a paid Copilot tier plausible, so Priya Shah’s revenue commitment depends on them. Viability — passes, conditionally. Each individual step is a well-bounded operation against an API; this is not a task requiring reasoning the model cannot do. Cost of error — fails as stated. A wrong credit is money out. A wrong reschedule is a technician sent to the wrong site. The 2023 routing feature degraded quietly and was switched off, and Ivan Petrov will read any proposal through that lens.

So the answer is not “no”, it is “not until the fourth criterion is engineered.” The remaining topics in this section do exactly that: bounds on the loop, a tool surface the harness can gate, enforcement outside the model, confirmation for anything that moves money, idempotency so a retry is not a second credit, and an audit trail that lets someone reconstruct a decision afterwards.

With those in place, the fourth criterion holds — errors are caught before they commit, or are reversible, or are visible. Without them, the honest recommendation is a chain that prepares the action and hands it to one of Grace Adeyemi’s 45 agents to approve, which captures much of the value at a fraction of the risk and would be a perfectly respectable first release.

Feeds the workbook

The rubric requires the agent-versus-pipeline decision to be argued with the simpler alternative genuinely considered. Write the five-line pipeline version in your agent design and state what it cannot do. That paragraph is the argument.

→ Agency granted, conditionally. The next question is what shape the loop takes. — Agent patterns