# Verification lets coding agents work for hours

Markdown copy of https://realalpha.app/s/verification-lets-coding-agents-work-for-hours/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Amjad Masad](https://realalpha.app/builders/amjad-masad/) · Amjad Masad is founder and CEO of Replit. · Co-founded and leads Replit, a browser-based coding platform letting anyone write and deploy code.
- Host: [Reid Hoffman](https://realalpha.app/people/reid-hoffman/) · Co-founder, Manas AI · Co-founded LinkedIn; one of OpenAI's first investors; Greylock partner.
- Host: Aria Finger
- Show: Possible
- Episode: [Amjad Masad on vibe coding, AI agents, and the end of boilerplate](https://realalpha.app/v/xPDQXfS_vTE/), Jan 7, 2026
- Moment: 42:34 to 45:17 of the episode (3 min), playing on https://realalpha.app/s/verification-lets-coding-agents-work-for-hours/; the episode from there: https://realalpha.app/v/xPDQXfS_vTE/?t=2554

## Quote

> “Replit Agent 3 came out in September. We now boot up a browser for the agent, have the agent use it to **test the application** for you”
> — [Amjad Masad](https://realalpha.app/builders/amjad-masad/) at [43:29](https://realalpha.app/v/xPDQXfS_vTE/?t=2609) · Possible

## Takeaways

1. Agents need feedback to catch errors before they compound.
2. A browser lets the agent test the app it builds.
3. Separate agents can build, test, and review work.

## About this segment

Amjad Masad says Replit lengthened useful agent runs by giving agents ways to test and review their work. He describes runs growing from two minutes to hours, with handoffs between agents.

From “Amjad Masad on vibe coding, AI agents, and the end of boilerplate”, Jan 7, 2026. [Watch the full episode](https://realalpha.app/v/xPDQXfS_vTE/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[42:40](https://realalpha.app/v/xPDQXfS_vTE/?t=2560) that current agents have yeah what do you see as the kind of the longest unsupervised work that agents can currently do? I think it's more possible than our tools allow it to be. So I think state-of-the-art, LMs, especially Gemini has done a great job at a million contacts, million token contacts, but others have 200,000 or somewhere there. The easy thing is to fit all history inside the contacts, but that's not feasible.

[43:12](https://realalpha.app/v/xPDQXfS_vTE/?t=2592) There are few components. If you put them together, then you can run agents a lot longer. One component is verification. The work needs to be verified. It's a form of supervision. If the AI agent is going on a path and doesn't have feedback from the environment, from the work, from the human, for someone, then we're very likely straight off the path and do something really, like if you make one mistake, I'll just compound over time. So Replit Agent 3 came out in September. We now boot up a browser for the agent, have the agent use it to test the application for you and we have another agent that's adversarial that's like reviewing the code and giving you feedback and because of that we went from agent one ran for two minutes unsupervised after that it didn't work very well. Nation 2 is 20 minutes.

[44:00](https://realalpha.app/v/xPDQXfS_vTE/?t=2640) Agent 3, 200 minutes, we're seeing 200 minute runs that are actually really, really useful. There's still mistakes and all of that. Now we've introduced an autonomy selector inside Replit. So depending on your risk appetite, you can go all the way high autonomy. And some people run it for 10 hours. Now, I don't recommend it to everyone. Now, if you have a lot of disposable income, you should definitely do it.

[44:23](https://realalpha.app/v/xPDQXfS_vTE/?t=2663) But we have medium autonomy and it can run for 100 minutes, high autonomy, it can run for 200, 300 minutes. So I think it's possible. In terms of the context, so the verification is very important. So you pass from one agent to another. And so one agent does a unit work. Another agent gets a prompt and says, test the app. You don't have to give it the entire thing. So it's a multi-agent system. And then after the testing is done, pass back.

[44:49](https://realalpha.app/v/xPDQXfS_vTE/?t=2689) But you don't pass it back using all the existing contacts. You kind of summarize and give the next. It's almost like passing the baton to the next one. Again, ideally, we would put everything in the agent, but I'm not sure we're going to be there anytime soon. So I think this sort of relates back to what we were saying, how the difference between programmers and non-programmers is collapsing. Because if everyone has a Replit agent, you don't need to be a programmer.