Most people picture an AI agent as a smarter chatbot — you ask, it answers. That’s wrong. An agent is a loop, not a conversation.
The loop has four steps: observe, decide, execute, evaluate. The agent watches a signal, determines whether action is warranted, takes that action, then measures the result. Then it loops again. There is no human in the middle unless you put one there.
That distinction matters because it changes what you build, how you evaluate it, and what can go wrong.
The anatomy of one agent
Here is a concrete example: a copy-rewrite agent running on a product landing page.
Observe. The agent pulls GA4 data on a rolling 14-day window. It watches one metric: conversion rate on a specific page. When that rate drops more than 8% below the 30-day baseline, the observation threshold is crossed.
Decide. The agent checks whether the drop is statistically significant (minimum 200 sessions in the window) and whether a rewrite has already been attempted in the last 21 days. If both conditions pass, it decides a rewrite is warranted.
Execute. The agent generates a copy variant using a prompt that includes the current page copy, the conversion signal, and a persona brief. The variant goes through a critic gate — a second model pass that scores the output against a rubric: clarity, specificity, no banned phrases, word count within range. If the critic score is below threshold, the variant is discarded and the agent logs a failure without shipping anything.
Evaluate. The approved variant ships to 10% of traffic via a feature flag. The agent monitors conversion rate on that slice for seven days. If lift is ≥5% with statistical confidence, it scales to 100%. If lift is negative or flat, it rolls back automatically and logs the result.
The whole loop runs without a human touching it. The human reviews the weekly log, not each decision.
Why narrow agents outperform general ones
A general-purpose agent — one that can “do anything” — has a large failure surface. When it fails, you often cannot tell why, or even that it has failed. The output looks plausible. The damage is quiet.
A narrow agent fails small and fails obviously. The copy-rewrite agent above either ships a variant or it doesn’t. The metric either moves or it doesn’t. There is no ambiguity.
Narrow scope also makes evaluation tractable. You can measure whether a single-function agent is working because it has one job with one measurable outcome. You cannot measure a general agent in any meaningful way — its success criteria are too diffuse to operationalize.
This is why every agent we run at Avakata has a single defined function. We have 160+ agents. None of them do “everything.”
The six agents every solopreneur should consider first
- Content drafting agent. Watches your editorial calendar and keyword gaps; drafts a structured post outline and first section when a gap is 14+ days overdue.
- SEO/GEO audit agent. Monitors your top 20 pages for title tag drift, missing FAQ schema, and citation-readiness signals; flags pages that fall below threshold weekly.
- Lead qualification agent. Watches inbound form submissions; scores each lead against your ICP criteria and routes high-fit leads to your CRM with a one-paragraph summary.
- Support triage agent. Reads incoming support emails; categorizes by issue type, drafts a response for common queries, and escalates edge cases with a summary.
- Social scheduling agent. Watches your published posts and repurposes key claims into platform-specific copy; queues posts for review on a set schedule.
- Invoice/bookkeeping summary agent. Pulls transaction data weekly; generates a plain-language summary of revenue, outstanding invoices, and anomalies for your review.
Each of these has a clear input signal, a defined action, and a measurable output. That is the pattern.
What makes an agent safe to run unsupervised
Three requirements. All three. Not two.
A clear success metric. The agent must have a single number it is trying to move. If you cannot state the metric before you build the agent, you are not ready to build it.
A critic or evaluation gate. Every output must pass a quality check before it acts on the world. The critic can be a second model, a rule-based filter, or a human approval step — but it must exist. An agent without a critic will eventually ship something it should not.
A rollback path. Every action the agent takes must be reversible, or at minimum, bounded. The copy-rewrite agent ships to 10% first. The support agent drafts, it does not send. If you cannot define what rollback looks like, the agent is not ready to run unsupervised.
Without all three, keep a human in the loop.
We publish the exact prompts and agent configs we use at Avakata. Subscribe to Field Notes at avakata.agency/contact.html to get them.
How to start with one agent this week
Four steps. Do them in order.
- Pick one repetitive task you do at least weekly that has a clear input and a clear output. Not a vague task — a specific one. “Write a first draft of the weekly performance summary” is specific. “Help with marketing” is not.
- Define the success metric before you write a single line of prompt. What does good output look like, and how will you measure it? If you cannot answer this, go back to step one.
- Build the simplest possible loop. One input, one action, one output. No branching logic yet. Get the basic loop working before you add complexity.
- Run it on a small slice. One page, one email thread, one week of data. Measure the output against your metric. Adjust. Then scale.
The goal in week one is not a production agent. It is a working loop you understand well enough to evaluate.
If you want to map out which agents make sense for your specific setup, book a discovery call. We will tell you what we would build first and why.