Building Agentic Apps End-to-End with Replit Agent

Framing: who's speaking and what this session actually is (~02:20–02:22)

Brandon Middleton is Replit's Head of Education. Two facts he leads with that aren't on a résumé: eleven years ago he entered the Guinness Book of World Records as the first person to rap a commencement speech (Haas evening and weekend MBA, class of 2015), and across jobs at Amazon (six years at AWS) and Microsoft he always taught on the side — Replit is the first place he gets to do it full time. His remit now runs from universities and colleges to K–12 and nonprofits, all aimed at making AI literacy practical and tactical.

He frames the session as both a talk and a demonstration: talk about education while building an app with Replit. A show of hands revealed a room heavy on technologists but with a substantial contingent of self-described non-technical people who vibe code, and he promised to serve both.

The first demo is an app he built that morning with Replit Agent: the room opens a URL, uploads photos and messages, and a "scroll wall" fills the big screen live. That same app becomes the running example later for databases, deployment, and APIs.

Thread 1: the premises college was designed on no longer hold (~02:24–02:30)

He teaches Redesigning Finance at the Stanford d.school every spring quarter, where students joke to him about the state of higher education — the tuition price against the classroom experience, and whether being in class is differentiated from going to YouTube. He thinks the jokes land because they're partly true: the belief that "I believe in the American dream, I did everything people told me to do, and it gets me a job" is starting to wane.

His central argument: the education system as it exists was actually designed pretty well — for a world that no longer exists.

• Information used to be scarce. You came to a university to receive it from a professor; knowledge was locked in libraries, laboratories, and lecture halls.
• Remembering the answer used to be the evidence of learning. Not anymore — you can have a lot of book knowledge and no street smarts, or a lot of observations that turn out to be misinformation or disinformation.
• Today information is everywhere, answers are abundant, and intelligence itself is becoming pervasive and abundant as model costs drop and model quality climbs.

Which produces the uncomfortable question for the institution: if we were designing college for today — summer of 2026 — from a blank sheet of paper, how would we do it? Would we still organize everything around lectures and libraries? Still run 16-week semesters, or make them shorter? Still count attendance? Keep closed-book recall and rote memorization on final exams, or throw them away?

His diagnosis is three confusions: we've confused time served with learning, credentials with capability to perform, and — sometimes — sorting students over four or five years with educating them. The result is an extraordinarily expensive sorting system: sorted by admissions, by grades and departments, sometimes by the institution attended — and the output four or five years later gets called merit, and people literally get hired or not hired on it.

Thread 2: the AI literacy tour and the signals from the field (~02:26–02:33)

Replit runs what he calls an AI literacy tour, which has been through Chicago, New York, Los Angeles, Atlanta, and the Bay Area, asking from the ground up: what does it feel like to be a student right now, and what does it feel like to be a teacher receiving AI mandates from above and trying to translate AI literacy into something meaningful for yourself and your community? The people he talks to range from preschool teachers to R1 researchers and professors.

One slide shows a community session in East Palo Alto with participants from age 7 to 77 discussing resource allocation issues in the city. He captured what the seven- and eight-year-olds said alongside what the 75- and 76-year-olds said, on sticky notes and his phone, then uploaded it into a Replit app and thought with AI as a partner — the point being to show a community that doesn't identify as computer scientists or engineers how easy it is.

The signals from the work world are stark:

• At the beginning of 2026, unemployment among young recent college graduates was about 5.7%.
• More strikingly, about 41.5% were underemployed — "I might have graduated from a really fancy school, but I might be working in something underneath my intellectual capability or even my interest."
• The World Economic Forum reports nearly 40% of workers' current skills are expected to change between 2026 and 2030. He translates it: if 100 people represented all eight billion, 59 of them would have to shift, reskill, and repivot within four years.

None of that means the institution has no value, he says — it means the connection between education, student capability, and opportunity is currently under strain and needs solving.

His own path is the argument: EE at UIUC → networking software engineer at Cisco in the Bay Area → Haas MBA in 2015 → putting down a lot of the programming and product management skills → and, after leaving AWS for Replit, picking those design/engineering/product skills back up to build for technical and non-technical people alike. Hence: learning, unlearning, and relearning is the core skill.

Thread 3: AI is not the next edtech gadget (~02:33–02:40)

He deliberately distinguishes AI from the calculator and from Google circa 2003. Yes, it automates grading, paper summarization, and lesson planning — but he sees a once-in-a-century opportunity to rebuild education around human potential. He believes everyone has an innate passion and skill to deliver something to the other eight billion of us, and that vibe coding and agentic AI can help people reach peak performance. "For the first time in history, imagining agentic assistants and personal tutors is a real thing that can be achieved within the next couple of years."

He is pointed about the distributional angle. In Palo Alto, neighbors send kids to Saturday school and to tutors that keep them out until 6pm — "students whose parents know how to navigate the system, that's pretty soft." What he cares about is the other side: students not confident enough to raise a hand in a crowded lecture hall. A 24/7 personal tutor means the student too shy to ask at 2 a.m. gets an answer; students who need an explanation a different way get one (visual versus auditory learners, students who do well in orals but not timed written exams); students who need accommodations get them without embarrassment, exhaustion, or judgment from a teacher who is often in a one-to-many setting. And multilingual support: living in the Bay Area for 21 years as a monolingual English speaker, he asked how many in the room speak more than one language — "so I am underachieving all of you" — but notes you can speak to Replit in French, Spanish, or any native tongue and have the apps you build translate for whoever is on the other side of the interaction.

He explicitly rejects the replacement framing. This is about expanding the reach of one person for a class of 15, 50, or 100. When the teacher becomes less a broadcaster and more a designer, coach, mentor, critic, and community builder, students get more room to grow into leadership and ownership. "AI can definitely deliver the information, but what the teacher needs to do — and why that matters — is that very personal care and concern, that intangible thing, that belief the teacher has in the student" that gets a student to stay up two extra hours or take pride in the work they turn in.

He stretches the boundary of "education" with an atypical case: a Grammy-winning musician, signed to a label for ten years and three albums, released from that deal five years ago, who came to the Replit office. His question wasn't technical — he wants to run his own record label and fix what plagued him inside the larger music industry, starting with revenue recognition for streaming royalties (fractions of pennies per stream across Spotify, YouTube Music, and the rest) and the runaround he got when he tried to audit what he was owed. No computer science background; he's using vibe coding to reimagine that part of the industry. "Even though he's not part of an actual university, that still falls under education in my scope — music education, business education, technology education."

Thread 4: so how does assessment change? (~02:48–02:52)

His position is blunt: "I think testing is pretty dead at this point."

In Redesigning Finance he asks students to record five- to seven-minute YouTube explanations of themselves grappling with a topic. One theme from last year's class: if you're redesigning reinsurance and the insurance industry to improve the claimant's experience for the next LA wildfire or the next Florida flood, how do you design at the claimant level, the insurer level, and the reinsurer level? Students build something and then defend it synchronously — "so unless they've learned how to master deep faking, they're putting the work into actually thinking through it and putting a prototype and their video next to it as a submission."

His grading is redesigned too: all group work, but adaptive — weighted by the student's year in school (freshman through master's) and by background (engineering versus brand-new to technology), and scored on the fidelity of the design plus collaboration as measured by how their teammates vote, not by how the instructors and TAs read the group.

He replaces the assessment question entirely. Not "did the student use AI?" but:

• Did the student actually think?
• Did the student verify the work?
• Did they cite sources and say where it came from?
• Did the student improve the work?
• Did the student explain the decisions and choices that led to the output?
• Can the student defend the conclusion?
• Can the student tell us where the machine got it wrong? — using their own judgment and expertise to evaluate what the machine said.

In practice: an oral defense after every major individual or group project, and students submitting not just a final answer but a tracked record of their reasoning. He built a system inside Replit for students to reflect on lectures, submit assignments, and track that arc — "the space between two lectures, you might have learned a lot. The space between one exam and the next, you might have learned a lot. And the traditional system doesn't capture those moments."

Two cases where process beat domain expertise:

• A San Francisco life-sciences hackathon where students used Replit for biomarker discovery and drug-discovery work. He admits he didn't do particularly well in molecular and cellular biology or chemistry, but as a judge his feedback focused on how deeply they thought about the solution and how many people they surveyed — their process — rather than his domain expertise. Even without going deep in a domain, lived experience and perspective can make what students build better.
• Jack, 15, at Holy Trinity High School in Chicago, who commutes 25–30 miles to school. The principal gave students Replit access last spring, and Jack took the school's own disciplines and principles and built a system teachers use from their phones to reward behavior aligned with those values. Students accumulate XP, shown on a leaderboard on the digital signage in the lobby as students walk in from the buses, redeemable for special things at the school. A vibe-coding platform being used to change the culture and environment of a school. Asked how he thought about learning AI, Jack broke it into three: learning with AI (tutor, collaborator, critic, translator); learning about AI (security, governance, how models and algorithms work — from the school's one teacher who did a deep professional-development dive and brought it back to the sophomores, juniors, and seniors); and learning beyond AI (Bluetooth streaming his app to the signage system, plus the community impact and issues around the school). Brandon sees those three as a good starting frame, extending well beyond the classroom into medicine, justice, and resourcing.

He's also candid that this becomes a philosophical and ideological conversation quickly: there are things technology cannot cure — economic inequality, poverty, war — where money is sometimes a factor but never the whole solution. He wants to talk to communities and students about those principles alongside the AI literacy.

Thread 5: the Replit platform tour (~02:42–02:48, 03:04–03:12)

He asked an audience member to define Replit for someone who's never used it, and got: "from idea to built AI application without the requirement of knowing much about the technology underneath." He took that.

What he actually clicked through:

• The prompt box, familiar from ChatGPT or Claude, next to an import button for work started elsewhere.
• 511 integrations and connectors, personal and enterprise — Dropbox, Discord, Slack, Google Drive, Databricks, Google Sheets. "It's a lot more fun to connect into all of these applications before you build the thing you want to build": go grab my Google Drive, my Dropbox, my Databricks and my Google Sheets, and make an application that does XYZ.
• Output types — website, mobile app, design, games, slide deck — which steer the model. The slide deck for this talk was built in Replit.
• Learn and documentation in the bottom left, with video content taking people from zero to 100 and a bit beyond.
• Modes: Lite for basic asks ("change my slide background from black to orange," "change the text of this slide"), Economy as the default middle-of-the-road reasoning, and Power for genuinely ambitious work — with the ability to click into each and select closed or open models. "Choosing the right model for the right thing you're asking Replit to do is part of the experience."

The last stretch is a vibe coding 101 walkthrough of building end to end:

1. Multiplayer. An invite button at the top of the workspace lets other developers prompt the same body of code — "in the same way you can build out a Google Doc with five or six people, you can build software."
2. Your first prompt. He starts newcomers on a personal portfolio page: "make me something that'll help me express what I've accomplished in my career," throw in a picture of yourself, go.
3. Publish, which abstracts hosting, database provisioning, security scanning — everything needed to go from localhost on your laptop to a URL that works for everyone.
4. Add a database. "When you uploaded your pictures — if it was just a front-end page I couldn't have done that." He opens Tools → Database live, showing Replit's development database (for pre-publish testing) alongside the production database: 22 rows in dev versus 32–33 in production. "That's the difference between doing something locally before you publish, and publishing it live to the world."
5. Add APIs — weather, stock time series, Shopify, or any service — so the app pulls fresh data into the interface on load.
6. Add auth. "The shortest line in this whole presentation that will allow you to do something cool is: add Replit Auth to my app." Five or six words produces a login portal with Google, Apple, and the usual providers, so users can log their sessions, pictures, and text.

He closes on the limitations of vibe coding, and is honest about them:

• You do need to know a little about how product building works to figure this stuff out.
• You'll hit unfamiliar errors. Screenshot them and drag them into the interface: "I don't know what this is and how to fix it." The same Replit Agent building your application is the one you ask to get unblocked — you don't have to wait until a human can help you.
• Complexity and technical debt. He thinks debugging is the most useful skill on the list, and learning it as a non-technical person is a lot of trial and error — time spent with the agent picking up what traditional engineers learn in class.
• Guardrails. Security scans before publish, plus interface-level validations and judgment about what's appropriate. "We do as much as we can on the platform side to protect users and builders, but there's some common sense and responsibility on your side too."

Thread 6: five commitments and proof of work (~03:00–03:04, 03:09–03:10)

Five commitments for teachers, researchers, and students:

1. Guarantee AI access for every student — from well-resourced universities down to what he thinks is an appropriate floor at middle and high school.
2. Redesign assessment around critical thinking rather than right-answer/wrong-answer. "There's a lot of gray area in the life we live today" — a spectrum is worth considering.
3. Reward teaching innovation as seriously as research. He sees a stark contrast in academia between rewarding good teaching and rewarding research, and wants higher ed shining a light on people who can coach 25, 50, or 100 at a time into AI literacy. "That's a big opportunity for us."
4. Replace the one-time degree with lifelong learning. He references an on-ramp/off-ramp idea discussed among Stanford adjuncts a few years back: once you start at an institution you're there for life — leave at 25 and come back at 30, or leave at 45 and come back at 47. He concedes the economics and the exact implementation are still open questions, but the human need for community is already understood.
5. Measure capability and mobility, not rankings. Giving students a proof of work system as they graduate and look for jobs is a big opportunity. "I thought in jest it'd be fun to have a funeral service for the resume, for the transcript, maybe for the GPA" — because where he thinks this lands is that you point me to your portfolio of built work and the evidence of everything you've done solo and collaboratively; and sitting across the interview table, your ability to tell the story, to show the forks you faced and the ones you chose and why, is what decides whether he's excited to bring you onto the team.

The Penn State story anchors that fifth point. Four recently graduated students came to the Replit offices, still building the businesses they started in college. He asked whether each of them ran businesses and attended classes. "Kind of" — for every class, one of them was the leader, went to that class, learned everything, and came back and taught the other three. Four classes a semester, each of them leading one. What did they do with the freed-up time from the other three? Built businesses — four of them over four years. Did they make money? "Five figures? A couple grand?" — "No, each of us made like a million dollars each over four years in ARR."

Which made him wonder whether, instead of GPA, transcript, and resume, the goal should be learning information, monetizing it, participating in the economy across four years, and walking across the stage with zero debt. He polled the room on whether that's reasonable, and added: "I think I would hire those people."

He also showed a photo of his three kids — 14, 10, and 7 — whom he calls his three startups, emptying his wallet at the end of every pay cycle to make them happy. His motivation is that in four years the oldest startup goes to college, and nobody knows what learning, education, credentialing, college, or work will look like in 2030. "If you have ideas about this and you feel like you know what's going to happen in 2030, please see me immediately after this presentation." He puts the experimentation window at about a year to eighteen months, maybe two years, to try everything and see what sticks.

Replit's Faculty Fellows program is the mechanism: a cohort now about 75 teachers large, with a monthly webinar, where people who said yes to experimentation build experiences on Replit, put them in front of students or use them in research, and report back on what worked and what didn't — so the next person who hasn't dipped a toe in feels more confident on the back of that shared experience.

An atypical teaching case: teaching Shaquille O'Neal to vibe code (~02:56–03:00)

In December, Brandon and Replit founder Amjad were on set in Atlanta teaching Shaquille O'Neal to vibe code. The first obstacle was physical: his hands are too big to type. The workaround was to open multiple browser tabs, hit F5, and just start talking.

In 90 minutes they built three things:

• Shaq GPT, to answer every sports question, with "Charles Barkley sucks" hardcoded into any question about Charles Barkley, his TNT broadcast partner.
• Shaq Daddy: "Brandon, I want to solve the loneliness epidemic in America. There's a lot of single men that need to know how to talk to women, and we need to get these men married" — a pickup line generator.
• An advertisement and investment tracker, since Shaq's brand is everywhere (he and Snoop Dogg are "neck and neck"), to see year-to-date 2026 or all-time returns on his time and his brand.

The apps aren't really the point. What he wanted from it was seeing someone like that be vulnerable and okay with not knowing things, and try anyway. That's the vision for Replit's education team: find people with the willingness and humility to try, and once they've done it, have them take it back to their audiences and their alma maters (Shaq went to LSU). "If you have ideas about how to scale this on a platform or a person that other people look up to, or that kids see as cool, that's actually part of my strategy on the education team." Expect more of it with Shaq and others as back-to-school season arrives.

Quotes

"We've confused time served in college institutions with learning. We've confused the credentials with a student's capability to actually perform. And sometimes we've confused sorting students throughout four or five years with actually educating them." (~02:29)

The whole indictment of higher education in three sentences.

"College has become this extraordinarily expensive sorting system." (~02:29)

The conclusion that follows from it.

"The question should no longer be, did the student use AI? … Did the student actually think? Did the student verify the work? … Can the student tell us where the machine got it wrong?" (~02:51)

The assessment question replaced wholesale — the last item being the real test of judgment.

"The same tool that you're using to build out your application is the same Replit agent that you can ask questions to in order to get unblocked." (~03:12)

For a non-technical builder, this is the most practically useful line in the platform tour.

"Each of us made like a million dollars each over four years in ARR for the businesses that we started." (~03:10)

The Penn State students' own words, and the strongest support for replacing transcripts with portfolios.