Office Hours #24 · Advanced session
Super AgentsDeep Dive
“Like a human employee in a chair.”— Sam, defining Super Agents
A deep advanced session on what Super Agents are, how to use them, and the new build capability that lets Super Agents send prompts to your app autonomously. Sam demoed entity querying, backend function calls, and fully autonomous feature building inside a live GrowthHub app.
- Query entities
- Call functions
- Build features
- Creative initiative
The operations map — one core, four docked capabilities, one linked app.
Announcements
Seven dispatches from the session.
The briefing
An employee in a chair,
not a bot in a box.
ELI5
Super Agents are autonomous employees for your apps. They can read data, call functions, run tasks, and now build features or whole apps by sending prompts to the builder themselves.
What Is A Super Agent?
- Sam framed Super Agents as “like a human employee in a chair.”
- You hire them with a blank canvas, give them skills, set guardrails, and let them execute.
- They can reference LLMs, interface with your app, your data, your functions, and external systems.
- Anything a human can do with a keyboard and mouse is in scope. Physical work is not.
Agent vs Super Agent
- In-app Agents are FAQ bots and knowledge-base interfaces that use integration credits.
- Super Agents are autonomous employees that use prompt credits, plus integration credits when calling functions.
- Agents are more rigid; Super Agents have autonomous reasoning and creative latitude.
- Agents live inside your app. Super Agents live at app.base44.com/supagents.
Hooking Up A Super Agent
app.base44.com/supagentsThe capability orbit
Four satellites, docked and demoed live.
Sam ran all four against the linked GrowthHub app. Focus a satellite to read what it did on the day.
Capability 1 — Query Entity Data
- The Super Agent can read linked app entities, summarize data, score items, prioritize feedback, and return a structured report.
- This does not require an app-side function when the data is already in entities.
- Demo prompt“Please summarize this week’s top ambassadors. Score or flag high priority feedback items.”
The back office
Skills, memory, tasks, credits.
Everything around the orbit that keeps the agent running: what it knows, what it remembers, what it does on schedule, and what that costs.
Building Skills For Your Super Agent
Just describe the expertise you want in chat: UX strategist, data analyst, QA tester, SEO writer, and so on.
Do not write skill files by hand. Use Claude or GPT to draft soul.md from your app context, then paste it back into the agent.
Give the agent tools, context, constraints, and a role — like training a human employee.
Persistent Memory Pattern
- Do not rely on context window memory for critical instructions.
- Create a hidden agent_instructions entity and store the rules there.
- In the Super Agent skill, reference the entity as the permanent source of truth.
- Context can compact. Entities do not.
Tasks Tab
Credits
Creating New Apps From A Super Agent
- Sam tested this live with: “Please create me a beautiful new app that shows me all the top 10 trending MMO RPGs worldwide.”
- It worked. Super Agents can spawn new Base44 projects.
- Sam also re-emphasized: always publish immediately after the first prompt, set private if needed, and test properly across devices.
Community Q&A
51 questions off the wire.
Everything the chat asked about Super Agents, answered live — from phone numbers and prompt credits to Playwright QA and whether Lily is real.
51 questions answered live
Absolutely. You could even build a Super Agent version of Sam by analyzing every Discord message, voice, and video he has done.
Yes. Every prompt the agent sends to your app has the same revert button as your own prompts. All tools have stop buttons, and you can delete the agent entirely.
Only if you give it a tool to check periodically, such as reviewing prompt history every 30 minutes. It does not observe in real time by default.
Yes — joke. Zlatan and Fred are just the running naming memes.
They both consume credits. In-app agents use integration credits. Super Agents use prompt credits and also integration credits if they call functions that use them.
Yes. Anything happening in the Super Agent chat window costs credits.
Type the expertise into chat, hook it to connectors, sync external workflows, give it tools, and add persistent memory through entities. If a human employee can be taught the task, so can a Super Agent.
Super Agents: what they are, how to use them, new features rolling out, and the headline that Super Agents can build apps.
Yes. If a human can do it with a keyboard and no physical work, a Super Agent can.
Sam was not sure and needs to verify. His guess was yes, but not confirmed.
Disconnect and reconnect.
Personal preference. Sam prefers individual agents per role — SEO, design, QA — but one agent can hold multiple skills.
Claude is only prompt engineering. The actual Super Agent lives inside Base44.
Technically yes via external API. Better practice: build image generation as a function in your app, then have the Super Agent call it.
Yes. Scheduling, looping, learning, and improving are core Super Agent strengths.
Yes, different numbers per agent.
Probably yes through a voice-to-text interface, but Sam was not fully sure of the architecture.
Not out of the box. Possible through complex APIs such as ElevenLabs plus text-to-AI, but not native.
Yes. Super Agents can build and manage entities entirely.
Sometimes Super Agents break. Build a new one in a fresh context window if the agent is stuck.
Yes.
app.base44.com/supagents or base44.com/supagents. Just talk to it.
No. It is not a data scraper. It can access apps you own.
Implement a scheduled task that backs up wherever you want, or export manually off platform.
Make the roadmap a database entity with statuses like pending, in-progress, and done. Have the agent reference it.
Ask to implement a third-party LLM for invoke calls and provide the API key.
Yes.
Yes. It can review your app and recommend improvements.
Yes.
Immediately after the first prompt. Set the app private if needed. Publishing is required for proper device testing.
Yes.
That is exactly the new build capability demoed today. Rolling out within 7 days.
Super Agents are more capable and reason more deeply, but cost more.
Yes. Sam tested it live and it worked.
Yes. Tell it: “We are rigid now. Do not update.”
Use Claude or GPT to draft from app context, then paste it back into the agent.
Yes. It is only as good as you make it, but the architecture supports it.
Yes. Sam demoed scheduling the weekly digest function every day at 3am UTC+1.
Yes. It spins up a browser and interfaces with it.
Build the API call as a function inside your Base44 app, then have the Super Agent call that function. It bypasses the loop.
One per app. Each app has its own user journey and edge cases.
Rename via chat: “your name is X.” The UI may not update the same way.
Depends on purpose. If the Super Agent connection works, leave it.
Build capability is not out for everyone yet. Sam has early access. Rolling out within about 7 days.
No. Paste the builder URL with the app ID. One message connects it.
Same as the builder: credit cost depends on output, not a flat 1 credit.
No hard limit. It stores infinitely, then compacts when context fills, keeping summaries but possibly losing exact detail.
Yes. AJ’s archive.
It already was on YouTube; Gary had a setup issue.
Sam joked: “Now we’re talking.” No comment.
No comment.
The recording
Watch the agent work.
The full session as it streamed live — every capability run against GrowthHub, the autonomous build reveal, the MMO-RPG app spawned on the spot, and all 51 questions from the chat.
Play recordingThe debrief
Ten things to file before you leave orbit.
Heard on the wire
“Like a human employee in a chair.”
“If a human can do it with a keyboard and mouse, a Super Agent can.”
“Please go and show your love to those developers.”
“Always publish immediately.”