Office Hours · Session #32
Super AgentsBuild With Us
The session chalked up like a training play: draw it on the board, then let it run off the pitch as a pipeline.
A deep, two-part Super Agents session. Part one: Sam walked through his existing Sam's Pizza Business Intelligence super agent — a fully autonomous BI suite running scheduled tasks for competitor research, weekly marketing, trend radar, and review responses, all for ~0.9 credits/week. Part two: live build — Sam took an idea from chat (Nick's youth soccer coaching app), built Pitch Perfect Training in two prompts, then live-built a super agent on top of it that would generate weekly training plans for all players. WhatsApp output integration went sideways when an audience member accidentally scanned the connection QR code on-stream and connected their own phone instead. Recurring theme of the session: building super agents is easy — the hard part is thinking creatively about what they should do for your specific situation.
Before kickoff
Team news
6 notices pinned up during the session.
Super Agents have a new "build" capability
Currently in staff testing only. Demoed live: Sam had his super agent receive a casual idea ("build me an app to sell my dope arcade machines") and the super agent autonomously drafted a structured Base44 builder prompt and dispatched the build (the resulting app was called "Neon Arcade Cathedral"). Coming to users eventually.
Plan Mode is coming to post-build chat
Confirmed again. Soon you'll be able to use plan mode on existing applications via Discuss Mode-like flow.
Amsterdam in-person event next Thursday
First Base44 community in-person event at the Tag offices in Amsterdam. Pizza, beer, credit giveaways, in-person Super Agents build. Sam is off Friday/Monday (Center Parcs in the UK) and flying out Wednesday.
Charles is hosting Thursday next week
While Sam is in Amsterdam. Drake & Robert hosting Wednesday again.
Sam's Pizza BI super agent costs ~0.9 credits per week
For the full autonomous BI suite (competitor research, marketing strategy, trend radar, review responses) — confirmed on the live cost preview.
Integration credits ≠ prompt credits
Super agents do NOT spend integration credits unless they call a backend function. Sam re-clarified this on a question from Andre about credit drain.
The training plan
The playbook
What Sam coached before, between, and after the two builds.
The team talk
The "human in a chair" reframe
Sam's repeated framing: "Your super agent is a little human in a chair where you instruct it how to do things. Imagine you're paying this person five grand a day. How do we get value out of that? They work 24 hours a day for you. What do they do?" The hard part isn't building the super agent — it's thinking through what you actually want it to do.
Drill 01
Super Agents are not "Agents"
The Agents inside Dashboard are essentially programmable chatbots — knowledge base lookups, scripted responses, basic data queries. Super Agents are something different: they're an army of employees you train to do whatever you want. Like having n8n/Zapier and OpenCore wrapped into one — but with full read/write access to your app's entities and the ability to be tasked autonomously on a schedule.
Drill 02
The duplicate-entry trap
Without the explicit instruction "on repeat runs UPDATE existing entries rather than creating duplicates," your super agent will add new rows every run — meaning the same competitor appears 5 times after 5 weeks. Always include the de-duplication instruction.
Drill 03
Cost vetting workflow
- Cost of any individual task: visible at task level when you press "Run now"
- Integration credits only spend on backend function calls — vet those functions for cost
- If you're worried about runaway cost, turn tasks off (archive/pause), then run manually to inspect cost per execution, then schedule based on that calibrated cost
Drill 04
Pending Mode for safety
For high-stakes actions: instead of letting the super agent act directly, have it move records to a
pendingstatus, then email/WhatsApp you for approval. On thumbs-up reaction, it proceeds. This middleman pattern is essential for anything sensitive.Drill 05
The plan mode / discuss mode equivalence
Sam re-confirmed: Discuss Mode and Plan Mode are now functionally Gemini-with-context inside your own app. You no longer need ChatGPT or Claude as an external thought partner — your builder has full context of your app and can theorise alongside you. "Long gone are those days."
The match plan
Two demos, two halves
First a finished play reviewed on the board, then a new one drawn up live in front of the crowd.
First half
Sam’s Pizza — the BI agent, replayed
The finished play: an autonomous business-intelligence suite running on a schedule.
The Super Agent Setup Pattern (Sam's actual prompt structure)
Sam walked through his real setup prompts for Sam's Pizza:
Prompt 1 — identity and scope:
"I will name you Sam's Pizza Business Agent. You'll be working on my app named Sam's Pizza Hub. You're full CRUD access. I've created a function in the app called
bulk_create_records. Please test that you can see and write to this app by creating a test record in the competitors entity."
Why the test write? Because by default, super agents tend to write to their OWN database (the super-agent app's entities) rather than the target app's entities. Force a test write early to confirm it's hitting the right app. Nine times out of ten Sam catches a mis-targeted write at this step.
Prompt 2 — the role definition:
"You are a business intelligence and operations agent for Sam's Pizza, a coal-fired pizzeria at 47 Mulberry, Little Italy, New York, NY 10013, established 1998. You have full access to the app's back-end... including 13 months of sales history, menu performance, game day records, marketing campaigns, customer inquiries, reviews, competitor intelligence, and trend radar entities. You assist me by answering business questions, drafting customer replies, researching and monitoring competitors, responding to reviews, and generating weekly marketing strategies. Always write in a warm, confident, local New York City tone. Never corporate, never generic, and always pull from real app data before responding. Quote actual numbers, pizza names, and dates where relevant."
Rigid, hard-set guardrails. Define what it is, what it does, the tone, and the data integrity rules.
The Four Scheduled Tasks (the autonomous BI suite)
Task 1 — Friday 4:00 PM — Competitor Research
"Research the pizza and fast food restaurants within half a mile of our address. Find five competitors. For each: address, current pricing, what they're actively promoting right now, their social media presence and approximate follower counts, their positioning, target audience, recent content campaigns or viral moments. Save each competitor as an entry in the competitor intelligence entity. Set the last_check date to today, and on repeat runs UPDATE existing entries rather than creating duplicates."
Task 2 — Friday 5:00 PM — Weekly Marketing Strategy
"Review the competitor intelligence entity for what competitors are currently doing. Based on that activity and this month's sales data, identify one promotional opportunity for us. Write three social posts — one for Instagram, Facebook, and X — all in a different tone and style. Save everything in a new weekly_marketing_strategy entry with this week's date and status set to draft."
Task 3 — Monday 9:00 AM — Trend Radar
"Search for two things: any food or restaurant trends gaining traction in New York City right now (social media, press, TikTok, local food blogs) and any major sporting or cultural events happening in NYC in the next 7 days. For each finding, write a recommended action for the pizza shop and a social post angle or menu addition. Reference past game-day records when suggesting pricing. Save each as a new trend_radar entry with this week's date."
Task 4 — Monday 8:00 AM — Review Response
"Check the reviews tracker for any entries with status pending. For each one: classify the sentiment, draft a response in our tone (warm and personal for positives, empathetic and solution-focused for negatives), and save the drafted response back to the entry. Flag any negative reviews from the past 7 days that remain unanswered as urgent."
Second half
Pitch Perfect Training — built live
The live play: from a chat brief to an app with its own super agent, in one sitting.
Live-build pattern (Pitch Perfect Training)
Sam built the coach training app in plan mode → 1 prompt for the build → 1 prompt for fabricated player/session data. Then created the super agent in 4 prompts:
- Name + app scope + test write
- Build out the role definition (the AI orchestra-style "thoughtful coach" prompt)
- Define the weekly task (generate session plans for all players, output to WhatsApp)
- Confirmation
Pinned in the dugout
The laminated rule
One card, laminated, taped up before every build.
Coach’s rule
The crucial "anti-write-to-wrong-app" instruction
Always include a test write to a specific entity in the target app on the very first prompt. Without it, you can do hours of work and find nothing in your target app because the super agent has been writing to its own database the whole time.
Play stopped
The whistle goes
Shouted in from the stands
Questions from the touchline
17 questions taken while the builds were on the board.
Showing 17 of 17 questions
The replay
Watch it back on the big screen
Both halves in full — the walkthrough, the live build, and the whistle.
Office Hours #32 — Super Agents — Build With UsYouTube · opens in a new tabFull time
The match report
Super agents are not agents. Different product. Super agents have full CRUD on your app and run autonomously like employees — agents are scripted chatbots.
The hard part isn't building — it's thinking. Sam's repeated framing: imagine a human in a chair earning £5K/day with 24/7 availability. What do they actually do? That mental model unlocks everything.
Always test-write to a target-app entity in your setup prompt. Otherwise your super agent writes to its own database, not your target app — and you'll find nothing has happened.
Always include "update existing entries rather than create duplicates" in scheduled task prompts. Otherwise you'll have 5 rows of the same competitor after 5 weeks.
Sam's full Sam's Pizza BI suite costs ~0.9 credits per week — autonomous competitor research, weekly marketing strategy with 3 social posts, daily trend radar, daily review responses.
Integration credits ≠ prompt credits. Super agents only spend integration credits on backend function calls — not on prompting. Vet your functions, not your prompts.
Use Pending Mode for high-stakes actions — agent moves records to pending status, WhatsApps/emails you for approval, proceeds on thumbs-up. Middleman pattern is essential for anything sensitive.
Super Agents will be able to BUILD apps soon (staff testing now). The Neon Arcade Cathedral demo showed what's coming: chat your idea casually, super agent drafts and dispatches a structured Base44 build prompt for you.
The kit bag
Everything handed out or pointed to during the session.
- unofficial-events.base44.app — AJ's community archive
- Apify — for LinkedIn / closed-platform scraping (paid, legality varies)
- AI Controls — bottom-left cog → custom instructions for persistent guardrails
- Testing Agent — three-dot menu next to Publish