Key Concepts
Scope creep (in agents): Not one bad decision, but building on the fly without a plan. Each addition seems reasonable and small, yet the agent you end up with is not the one you designed. Designing before you build is the fix.
Cut list: A written record of everything you removed from the first version. It turns “no” into “not yet” and tells you what to build second, so the next version starts from a decision you already made.
Quick Reference
The three-part scope
- One pattern: watching, finding out, creating, or reporting. Pick one, not two.
- One output: a single thing you can look at and say “that worked.”
- One cadence: a schedule you named, not whenever you think of it.
How to scope before you build
- Write down the output you want first
- Ask where the result needs to end up, who needs to see it, and in what format
- Work backwards from the output to the steps that produce it
- Confirm you have access to the data every step depends on
- Write down everything you cut
Warning signs your agent is too big
- Only the person who built it can follow it
- When it returns something odd, nobody can explain why
- The team would rather start over than fix it
- It started at a few steps and grew through small additions
- It depends on data you cannot access or cannot stand behind
Example: Mary’s competitive positioning agent
- First attempt: 22 steps, grew from three, relied on competitor data she did not have access to. It broke constantly and ran on data she could not stand behind.
- Rebuild: write down the desired output first, a map of where the brand sits against competitors in Answer Engines. Then work backwards: pull your own numbers for a baseline, pull competitor numbers per engine, compare them.
- Result: 30 minutes to rebuild, and it ran without failing after that
Benchmark: Aleph
- Constraint: small marketing team with little capacity to track AI citations, competing against well-established firms
- Approach: one theme per month, such as competitor-adjacent prompts or questions about specific software capabilities. One month’s theme was AI-powered variance detection, which became a buyer’s guide.
- Agents handled: the repeatable half of the content work, including building briefs, finding top-ranking outside sources, and choosing internal links
- Results: citation share rose from 1.5% to 5.3% in two months, and LLM-attributed traffic rose 82%
Notes
- Ship one small agent, then move to the next. The goal is a set of small agents working together, not one big agent only its builder can follow.