Key Concepts
Context: What an agent knows about your company that you did not write into its instructions, such as brand guidelines, messaging, and examples of past work. It is not one of the five parts. Attach it from a Knowledge Base so you can update your brand once without rewriting the agent. A new teammate who never got the binder will guess, and so will an agent.
Trigger: What decides when the agent runs. It is the difference between an agent you use and an agent that works for you.
Quick Reference
The five parts (think of a new teammate on day one)
- Instructions: the job description. What the job is, what a good result looks like, and what to do when something unexpected shows up. Most agent problems are instruction problems.
- Inputs: the references you hand over each run. The job stays the same and the input changes. Example: the topic for the Content Brief Generator.
- Tools: the apps and actions the agent is allowed to use. An agent can only do what its tools let it do.
- Triggers: when it runs. Someone asks, a schedule, or something changes.
- Artifacts: what the agent hands back, such as a brief, a report, or a message.
The five kinds of work every step falls into
- Retrieval: getting data that already exists, like your own metrics
- Research: going out to find something new, like what a competitor published
- Reasoning: thinking, which means summarizing, scoring, or drafting
- Logic: deciding what happens next based on what reasoning found
- Delivery: sending the result somewhere a person will see it
Diagnose by symptom
- Wrong number in the brief: retrieval
- Missed a competitor’s page: research
- Read the data and drew a poor conclusion: reasoning
- Sent a summary of a week where nothing happened: logic
- Nobody saw the output: delivery
Common failure to avoid
- Changing the wrong part, rerunning, and getting something different but no better. Name the failing layer first, then change only that part.
Reference build: AEO Content Brief Generator
- Takes a topic and turns it into a content brief grounded in what answer engines are already citing
Benchmark: Optro
- Ran optimization agents across multiple articles at once using Profound Sheets, then published directly to their CMS
- Updated 31 articles in one week
- This volume works only when each part is set up on purpose: the right instructions, the right context, and a clear place to deliver
Notes
- You do not need to build an agent by hand to use this. You need the names, because they are how you describe the work and how you diagnose it.