# Is this an agent problem? Three gates before you build (/docs/agent/is-this-an-agent-problem) --- type: Procedure title: "Is this an agent problem? Three gates before you build" description: "Before any agent is built, the three checks from The AI Agent Factory: does it need an agent at all, is it a once or an every-week task, and what does finished look like." status: stable order: 2 generated: { by: "process:claude-code", at: 2026-09-09T07:45:00Z } verified: - { by: "human:rrizwan98", at: 2026-09-03T08:20:00Z } - { by: "human:rrizwan98", at: 2026-09-09T07:47:26Z } sources: - id: af-28 title: "Is This an Agent Problem? A Crash Course. The AI Agent Factory, Panaversity, 2026" resource: https://agentfactory.panaversity.org/docs/is-this-an-agent-problem ksor: audience: [public] owner: human:rrizwan98 approval: { by: "human:rrizwan98", at: 2026-09-09T07:47:26Z } trust_tier: human-reviewed build_id: sha256:4d4bff09bd0a1dcff0a0a50f20f7ef514fc471e2b1f9e81e0e7b6150b93c0793 dirty: true --- Your step table from [count your own workflow](../method/count-your-own-workflow.md) says how much of a run could be automated. It does not say whether the run should become an agent at all. The AI Agent Factory puts three gates in front of that decision, and its rule is worth keeping in one line: **the cheapest mistake to fix is the one you catch before you start.** [^af-28] Run every candidate through the three gates, in order, before anything in this room is built. ## Gate 1: does this need an agent at all? Two checks, in order. [^af-28] **Check 1a, AI or a regular tool?** If the task is one exact step that is the same every time, a regular tool is faster and more reliable: adding up a column is a spreadsheet, finding a name is the search box, changing every "2024" to "2025" is find-and-replace. The task is an AI job only when it needs judgment, mixes different kinds of files, or no app exists for it. [^af-28] **Check 1b, an agent or just a chatbot?** A chatbot answers; an agent does. If you only need to know something, a draft or ideas, that is answer work. If the tool must open your files, change them, run something or use another app for you, that is action work, and only action work is an agent job. [^af-28] In the book's office example, two of four "AI tasks" turn out not to be agent tasks at all: a total is a spreadsheet, an explanation is a chatbot, and only the two that read and act on real files pass the gate. The gate is not there to push work toward the agent; it keeps work that does not belong there from getting there. [^af-28] ## Gate 2: once, or every week? This gate names the two kinds of agent work: [^af-28] - **Mode 1, solve it once.** Open an agent, do the task, take the result, walk away. Nothing stays behind. Most work, most of the time. - **Mode 2, build a helper.** A permanent AI worker that does the task again and again without you. The book calls it a Digital FTE. A task is Mode 2 only when all three dials are up: [^af-28] | Dial | Up means | | --- | --- | | How often? | Again and again, every week or every day | | How same? | Same kind of input, same steps, same finished result | | Worth it? | Big enough to repay building: time per run, items handled, cost of a mistake | If even one dial is down, stay in Mode 1. The expensive mistake is the quiet one: doing a Mode 2 task by hand, over and over, because no single run ever felt big enough to stop and decide. The book's trigger: **the third time you do the same task the same way, stop and run Gate 2.** [^af-28] Your step table already answers two dials. A run you counted is a run you repeat, and its steps are the "how same" evidence. The third dial, worth it, is the two numbers on the table: how much of the run a system could own. ## Gate 3: what does "finished" look like? Before opening an agent, write three lines: [^af-28] 1. **Works from:** exactly what the agent should look at. "My emails" is too vague; "the 400 messages in my Support folder from the last 7 days" is clear. 2. **Want at the end:** the thing that should exist when it is done, and its shape: a spreadsheet, a one-page summary, a folder of renamed files. 3. **Done when:** the one check that means the job is finished and correct, testable in under a minute. "Every message is in exactly one group, and the group counts add up to 400" is a check; "the summary is good" is not. These three lines are the target, not the instructions. They are also the seed of the spec on the next pages: the book's own crossing rule is that your brief becomes the spec, made explicit and permanent. ## Where each answer sends you The book's reference card, kept tool-free on purpose because tools change and the gate you land in does not: [^af-28] | Where the gates sent you | What you do | | --- | --- | | Just an answer (Gate 1, chatbot) | Ask a chat assistant; nothing of yours is touched | | Solve it once (Gate 1 agent, Gate 2 Mode 1) | Drive a coding agent in a session: it acts, you watch, you walk away | | Manufacture it (Gate 2 Mode 2) | Build a worker: this room, on the OpenAI Agents SDK | This room builds the third row, in its basic form: one text-only agent, one job, human gates kept. If your task lands anywhere else, you have saved yourself a build, which is the point of the gates. ## Run it on your own task Take the run you counted. Gate 1: regular tool, chatbot or agent? Gate 2: all three dials up? Gate 3: the three lines, with a done-check you could confirm in under a minute. If it comes out the far end, continue to [understand the work](understand-the-work.md), which reads the work itself before any tool is placed in it. [^af-28]: Is This an Agent Problem? A Crash Course. The AI Agent Factory, Panaversity, 2026.