Every agent I run on my businesses is built the exact same way: Context, Connections, Workflows, Memory. Here’s the full structure with real examples from the team already running my companies.
An Agent Is Not a Chatbot
A chatbot waits for a question and answers based on what it’s been trained on. An agent has a job. It has permanent context, connected data, repeatable processes, and memory that compounds. The difference shows up the first time you ask a chatbot a question that requires knowing your business and it gives you a generic answer that’s already in your last 50 LinkedIn ads.
My team of agents (a Meta media buyer, a Google Ads buyer, a Klaviyo strategist, a creative director, and a co-founder) all live in Claude Code. Each one is built with the same 4 parts. Once you see the structure, you can’t un-see it — and you can build your own.
Context is the set of files that teach the agent who you are, what your business sells, what matters, what benchmarks to use, and how decisions should be made. It’s not a one-line prompt. It’s a full reading list the agent goes through every time you work together.
In Mine
The 4 context files I give every agent
1. Identity file — who the agent is, what role it plays, what it refuses to do. (For my media buyer: “You are a Meta performance specialist. You optimize for profit, not vanity metrics. You never approve a budget increase without a written reason.”)
2. Business profile — what the brand sells, who buys it, margins, AOV, return rates, the 3 hero SKUs, the seasonality. The agent reads this every session so it doesn’t recommend things that don’t fit the model.
3. Benchmarks & rules — my CPA targets, ROAS thresholds, “kill an ad after 2x CPA with no purchases,” “scale at 3x ROAS for 3 days.” Plain-English business rules so its analysis matches mine.
4. Decision memory — key calls we’ve made before and why. So when we revisit a similar situation, the agent already knows the precedent.
The Mistake Most People Make
They write one giant prompt and call it a system. Context lives in files the agent re-reads every session — not a single message that scrolls out of view after a few exchanges. If the agent has to be reminded of your business every time, you don’t have an agent. You have a chatbot.
Connections are the live data feeds the agent pulls from. APIs, exports, scrapers, MCP connectors — whatever it takes to get the agent past “based on training data” and into “based on what your account is doing right now.”
In Mine
What my agents are plugged into
Store data: Shopify (revenue, orders, AOV by SKU, inventory).
Email: Klaviyo (revenue per recipient, flow performance, list growth).
Ads: Meta Ads (now via Claude’s native connector), Google Ads, plus Apify for scraping competitor ad libraries.
Customer voice: support tickets, reviews, Reddit and Twitter scrapes for category sentiment.
Industry signals: a curated list of expert accounts on X / Twitter that the agent scrapes weekly so I never miss an algorithm shift or a tactic worth testing.
Why This Is the Hardest Part
Most non-developers stop here because connections feel scary. They’re not. Most platforms have free or low-cost APIs, and Claude Code can write the integration code for you. The first agent I built took a weekend; every one after that took a few hours because the patterns repeat.
Workflows are the agent’s repeatable processes. The exact reports it builds. The exact audits it runs. The exact decision trees it walks through when something looks off.
Daily
Morning brief
Every agent has a morning routine. Pull yesterday’s data. Compare to the trailing 7 and 30 days. Flag anything outside expected ranges. Surface the top 3 actions with the reason for each. Done in one structured output.
Weekly
Account audit + creative brief
Once a week the agent runs a full account audit (winners ready to scale, losers ready to kill, frequency creep, conversion drift) and turns it into a brief my team can shoot from — specific hooks, specific angles, specific test hypotheses.
Monthly
Strategy review
A bigger pull-back: budget allocation across channels, what worked vs. what underperformed, what to test next month. The kind of work that used to require a strategist or an agency.
Workflows = Repeatability
A workflow isn’t just “a prompt I send sometimes.” It’s a written-down, named process the agent runs the same way every time. That’s what turns AI from a tool you reach for into a teammate you can trust to deliver on a schedule.
Memory is the part that makes the agent compound. Without it, every session starts from zero. With it, every session builds on the last one.
What Gets Saved
Decisions, outcomes, patterns, and feedback
Decisions: “On May 3, we killed the gym creative because frequency was 8.2 and CPA was 3.1x target.” The agent remembers, so when we see similar conditions, it doesn’t need to be re-walked through the logic.
Outcomes: “The static testimonial creative we tested in April hit 4.8 ROAS for 14 days, then dropped.” Pattern recognition for the next test cycle.
Feedback: when I push back on an agent’s recommendation, that pushback gets saved with the reason. Next time, the agent reasons differently. The behavior compounds.
Why This Is the Differentiator
Most AI tools you’ve used forget you the second the chat closes. An agent with memory gets smarter every week because every session deposits something. After a few months, you’re working with something that knows your business better than most contractors you’ve hired.
All four parts get built inside Claude Code. That’s where the agent’s brain lives — the context files, the connection scripts, the workflow prompts, the memory directory. You can run an agent fully manually from there: open Claude Code, ask the agent to run its morning brief, get the output.
OpenClaw is the optional second layer. Once an agent works manually, OpenClaw lets it run on a schedule and report back to me automatically — usually through Telegram. So instead of me asking the media buyer to run the morning brief, the brief shows up at 7am every day. Same agent, same workflows, just on autopilot.
You Don’t Need Both Day One
Build the agent in Claude Code first. Run it manually for a week or two so you trust the outputs. Then, if it’s saving you real time, plug it into OpenClaw to run on a schedule. Don’t automate something you don’t trust yet — that’s how agents drift.
For reference — here’s the full team I’ve already built using this exact 4-part structure:
Meta Media Buyer
Watches my account daily. Flags ads to kill, ads to scale, and creative ready to refresh.
Google Ads Buyer
Audits intent, protects Brand Search, and finds wasted spend in non-brand campaigns.
Klaviyo Strategist
Reviews flows, subject lines, segments, deliverability. Finds revenue we’re leaving on the table.
Creative Director
Pulls top vs. bottom creatives, mines customer language, builds production-ready briefs.
Co-Founder Agent
My strategic partner. Knows our businesses deeply, pushes back on weak ideas, keeps me focused.
+ Whatever’s Next
Same 4-part recipe. Whatever role I keep manually doing, it becomes the next agent on the team.
The Real Win
An agent isn’t magic and it doesn’t run your business alone. What it does is take the analysis and decision-prep work that used to require an analyst, an agency, or a strategist — and make it happen on a schedule, with your business context, every single day. That’s the leverage. Once you have one agent doing it, you’ll never go back.
For Your Job
If you’re ready to set up Claude for your specific job — with custom skills, connectors, and automations built around the work you do every day — I built a bootcamp just for you.
Start the Weekend Bootcamp →Go Even Further
If you’re looking to go even further, join mine and my husband’s community group where we give you all the AI agents and systems running our businesses.
Join the Community →© 2026 Mariah Brunner. All rights reserved.