← All AI Engineer talks

AI Engineer World's Fair 2025

Vibe Coding, with Confidence

About this talk

Qodo CEO and co-founder Itamar Friedman argues that enterprise vibe coding requires reliable, context-aware workflows spanning the software development lifecycle, rather than faster code generation alone. He introduces a newly launched command-line tool, discusses testing, review, bug fixing, and refactoring, and describes how MCP and A2A can support coordinated swarms of specialized agents.

Chapters

  1. 0:00Itamar Friedman introduces confident vibe coding and Qodo's CLI
  2. 1:36From IDE autocomplete to AI across the software development lifecycle
  3. 5:08Enterprise workflows, testing, refactoring, and agent interoperability
  4. 10:57Context-driven development workflows
  5. 18:18A2A coordination and the future swarm of specialized agents

Talk transcript

  1. 0:00

    [on-hold music] Okay.

  2. 0:15

    So my name is Itamar Friedman, the CEO and co-founder of Qodo, and I'm gonna talk about, yeah, vibe coding, but more on the confidence part and why, this is maybe surprising, why CLI might be the new interface leader for that.

  3. 0:29

    Uh, and, uh, we are soft launching a CLI tool, uh, right now, five minutes ago. So you can go, so you can go and install it. Uh, it's soft launch.

  4. 0:38

    Uh, play with it, give us feedback. Uh, there is some capacity. And, uh, in order to verify, like who is in the audience, just like if you can tell me if you tried Claude Code or Codex CLI.

  5. 0:50

    Okay. And I guess if I ask about AI IDEs, you would also, also do that, right? And the question is why, and I'm gonna talk about it a bit.

  6. 0:58

    Okay? Uh, so the agenda is that I'm actually going to do is really zoom out to the Gen 3.0. Uh, what do I mean by that? And from the autocomplete to the multi-agent workflow and workflows, you see it's a theme that re-returning here.

  7. 1:13

    CLI is something that can support that. Uh, and how do we believe at Qodo is gonna be the real shift in software development. And, uh, like, why is the CLI is the future in, you know, being able to do vibe coding with confidence.

  8. 1:28

    And eventually I'm gonna connect, do zoom out again. I'm gonna zoom in, zoom out, even with a demo, uh, and connect the dots. And I will, uh, say what...

  9. 1:36

    uh, explain a bit what is Qodo. I didn't do it right now, and we'll do it a bit, a bit later, but that's something I'm gonna cover shortly. So, uh, AI for developers, and I just think that most of people here are developers according to the amount of, uh, hands that were raised.

  10. 1:51

    So I think, like we first got some tools that we liked, AI tools, then we got some tools that we loved, and I'm not that sure that we actually, uh, get tools into our hands that completely, um, be-become a game changer for us.

  11. 2:06

    Okay. Uh, my-- it's maybe unpopular opinion. I'm gonna elaborate a bit. So in order to do that, I'm gonna differentiate developers between noobs and enterprise. It's a bit weird.

  12. 2:16

    Like, it's not necessarily juniors, like there are juniors in enterprise, right? But still, you know, uh, like let, let's, let's use that for, for a second. Noobs versus enterprise.

  13. 2:25

    So I think Gen 1.0, uh, was the autocomplete moment, right? We started, uh, being able to generate, uh, writing our, our, our code and getting a few lines ahead of us.

  14. 2:36

    We really liked that. If you weren't a developer, like y-you couldn't really do more, right? So you needed to be a developer already. And then came, uh, the Gen 2.0, the chat, the agentic chat, and I think like we saw even junior developers generating much more lines of code.

  15. 2:52

    Uh, maybe the seniors or the enterprise liked it, but also had a lot of problems with it. They need to review much more code, not necessarily high quality code.

  16. 3:01

    I hear that a lot, and I'm going to show you like, uh, five hundred tweets, uh, a-about it, not really, but, but from some, uh, opinion leaders. And I think that the game changer, uh, moment is when we will move from chat, agentic chat, to actually being able to, uh, and as a spoiler, command, command-line interface, to

  17. 3:19

    give commands to, to our agents as if they're like team members and they do end-to-end flow. What is an end-to-end flow? Is something I'm gonna talk about a bit.

  18. 3:29

    Uh, and I think like you will see different players in, in different parts of, of this map. Like, I mean, different tools. Let's, let's think about it again. So I think like in the, in the autocomplete moment, we saw first IDE plugins.

  19. 3:41

    Again, it's not a game changer. I'm just like going to complete more lines of code, uh, uh, for you. Uh, we liked it. And then, uh, let's rethink how we do coding inside IDE.

  20. 3:51

    So we saw IDE forks. And then what's the third wave? Sometimes I hear, uh, forking people. I didn't say that. What I think it actually we're gonna see is like i-- AI across the SDLC, okay?

  21. 4:03

    Like being able to do end-to-end poi-- end-to-end, uh, uh, flows that are specific, c-closing a task that is specific in the SDLC, but more importantly, using AI across the SDLC.

  22. 4:16

    Again, not in, not only in the IDE, hence how do we do that? Command line, right? It's one of th-those tools. Um, a lot of spoilers. So I think, uh, when we had the aha moment with the, uh, with the autocomplete, the, the next aha moment was, you know, all the hype around vibe coding, and I'm gonna

  23. 4:32

    touch about that. And I think the, the next, like thing that I think we're gonna hear about is like, "Hey, I'm really doing one hundred x," choose, choose a number, "ten x, fifty x, one hundred x development, and I-I'm doing that with confidence, not just like vibing through it."

  24. 4:48

    I think that, that will be a game changer, uh, uh, moment, uh, for, for us. It means that we totally broke the way, the way we work. We're gonna get to that.

  25. 4:56

    So I think like I, I differentiated between... I separated between noobs and enterprise, for example. I think for noobs it mean end-to-end means, for example, that you prompt and get a software, a fully managed software for you, right?

  26. 5:08

    We-- I, I guess we're all playing it with, even if you're a developer, you're doing some POVs or some site or being able to do some simple games. But then in the enterprise setup, in the heavy duty setup, then basically what means for us, like AI across the SDLC is reliable, versatile workflows.

  27. 5:25

    And when I say workflows, we had a few sessions, amazing sessions. I'm actually like combining them and which is one of, one of the points that were made there.

  28. 5:32

    Like, uh, uh, I, I think, uh, Charles makes, m-made a, made a point where an agent is creating workflows and calling them, and then there was a discussion about like workflows and, you know, uh, using agent, et cetera.

  29. 5:44

    So I-- when I say workflows, I already mean agentic workflow. Let's, let's, let's move on, like workflow, agent, who cares? Like, uh, like, uh, so it's, it's gonna be combined.

  30. 5:52

    And, and I think like in order to do that with trust, uh, y-you need to be able, more importantly, is to have workflows or agents that are related to trust.

  31. 6:02

    For example, having really high quality reviewing capabilities, really high quality testing capabilities, being able to do-

  32. 6:10

    A variety of jobs to be done that we do as developers in e-end-to-end-ish and high, and high quality. So that will be a game changer. Let's be b-bit more specific.

  33. 6:19

    As, as developers, uh, we do a lot of things, uh, not just write code. I took like four parts, which is like I think less than a half to n- of what we do, and, and I, and I think like basically vibe coding was very focused on let's like do initial planning and, and, and write code fast.

  34. 6:34

    Like, and, you know, vibe. W-whatever we get, like that implementation or that, we feel it and, and move on. But actually in heavy duty software, we need to do many other things.

  35. 6:45

    We need to do, for example, uh, bug fixing, or we need to do refactoring and changing features, et cetera. We wanna move fast with fault-breaking things. So I think we need to have not the, the testing and the writing as a, uh, a hand site, which you can do, like testing and, and reviewing inside IDE tools that

  36. 7:03

    are very mature, but not as a hand site, but at a... rather like really being the, the red team, right? Thinking as a, as a tech lead, as reviewing code and, and having like fitting technology, uh, uh, for that.

  37. 7:16

    So if we are able to, to put like AI in each one of those like parts, maybe then w- it will be a game changer. But I think even then, like we already have tools for, for code review, for testing.

  38. 7:29

    Shameless plug, we're gonna talk about later, that's what we do at Qodo. We focus being more on the red team. Even though we have tools for the blue, we are focused more on the red team.

  39. 7:36

    Still, I don't think it's a game changer, okay? And I, and I think like a first, uh, like a, a feeling of a game changer as I mentioned, not only putting AI across the SDLC, but having workflows across parts of the SDLC.

  40. 7:49

    So one semi game changer could be like if we manage to shift left the review and the testing. As, as you write code, then you also get those suggestion and get high quality code, uh, uh, that is already according to your company best practices, and you don't need to wait for the AI review tool to tell you

  41. 8:07

    that only after you pushed it, right? But I think like, uh, I think we, we have even a better way to do that, a bit m- more agentic. So let's an- reanalyze th- these four blocks, these four tasks.

  42. 8:20

    So, uh, I... instead of having one, like the x-axis, now there's also y-axis. It's executability. If you know the V shape in Wikipedia or whatever, where we're searching now, ChatGPT V shape software development, y-axis executability, uh, x-axis time.

  43. 8:35

    So we start with planning, it's not executable. Then we write code. In most cases you're not doing, admit it, like a TDD, okay? So you write, you write your code, and then you write your test.

  44. 8:45

    Uh, and m- most cases, and that's executable, but it's not application code, so it's a bit higher. And then, and then you do the review. A game changer is like when we actually manage to take that V and squeeze it and, and have like different workflows and agents talking to each other.

  45. 9:01

    And that's why like I had the word MCP and A2A, and I'm gonna get to that later. I, I, I, I'm g- I guess like if you're in this conference you know what I'm talking about, that where maybe there's different agent that are talking to each other.

  46. 9:11

    Maybe it's one agent using tools that are able to connect all of these like all, all at once. That, that will be a game changer. So to... and, and to some extent, what it means is that, that, uh, like the game changer moment is when we're looking all of these steps like rigorously and in one holistic, uh,

  47. 9:27

    uh, solution. If you like, uh, cloud is, is mature for 10 years, and then Wiz came and said, "We can't put an a- a security application for each part of the cloud, like, uh, or each, uh, uh, n- necessity of, of security.

  48. 9:41

    Let's build one holistic, uh, solution for, for, for cloud security." The same thing will happen to us in the game-changing moment for, for, uh, intelligence software development. Zooming out back again to vibe coding with confidence, uh, it, it didn't happen a long time ago when Karpathy, um, wrote about vibe coding.

  49. 10:00

    Uh, like y- really like I thought it's like a decade already in AI, uh, years. Uh, and, and the interesting thing, it didn't take too much time, like a month or two, uh, where, where we started, uh, where, where we started seeing pushbacks, right?

  50. 10:14

    This is not me, by the way. Uh, and, and, uh, and, and actually just a few weeks ago, the same Karpathy came back and rethought vibe coding. Notice what he's saying, like, "Code I actually and professionally care about contrast vibe code."

  51. 10:30

    The same Karpathy, okay? And it's really interesting I think to analyze what he's saying. The first thing he's saying, "Okay, cod- code that I care about, probably heavy duty software or I ha- like the quality, it's maintainable."

  52. 10:41

    But l- we are gonna get to that. And then he's, he tries to offer solution. I, uh, I love him. I, um, I mean, like I never met him but, but, uh, y- we all saw his videos, right?

  53. 10:49

    Like 15 years ago, 10 years ago. So, uh, he's giving solution, not only like, uh, like being annoying, like, "Okay, no, I said vibe coding, no vibe coding anymore."

  54. 10:57

    So first he's saying... the first number one is, is context. He's saying like, "Bring context." N- here he's saying doing it manual, but that's not game changer if now I need to start bringing manually.

  55. 11:08

    And then he's giving also two, three, four, five, six, seven. Like this, this is a workflow. He's actually suggesting a workflow. That's a- one of the topics of the three talks that we had now, okay?

  56. 11:19

    So now, uh, like this is really important I think, like how do you vibe code with confidence? You need it... y- like you said, with like just vibe coding, probably you don't care about qual- quality.

  57. 11:30

    You do care. You need to bring the right context and have workflows, and these workflows can be connected to agent, could be part of the agent, could be part of your flows, and that's what like in the nine minutes or so some, something I'm gonna talk about.

  58. 11:42

    Karpathy and others are not the only one. This is just like from a week ago, uh, if you, you know, this, uh, a very, uh, um, great, great person, great material.

  59. 11:51

    He's saying like, "AI coding tools are so much less useful on existing large code bases." Okay? He, he's saying that and, and this is differently than compared green f- greenfield work.

  60. 12:03

    And it's not surprising because if you have a greenfield work or it's a simple software, code generation is all you need, okay? Like it's magnificent, the tip of the iceberg.

  61. 12:14

    Give a prompt You get your software. But when you're talking about enterprise heavy duty software, then there's much more than that. We all know it, right? There is a maintainability, testing, reviewing, et cetera.

  62. 12:24

    And shameless plug, uh, well, I am the CEO of Qodo, I have to say, we are looking at that as a holistic, like how can we really bring a game changer moment as if we look on all of this at a holistic approach.

  63. 12:35

    That's why, uh, we have m- a multi-agent, uh, architecture. I am not going to go over it one by one because there's one specific thing that I wanted to show.

  64. 12:45

    So we have, uh, a tool... Uh, we have a agent that is focused on, let's call it deep research, deep ask in code. And, and then, uh, we have a tool that is specific, that's our flagship for, for code review called Qodo, Qodo Merge.

  65. 13:00

    And then finally, we have, uh, like a, a tool that is inside IDEs that enable to shift left, like bring that context, bring that, uh, bring that c- best, best practices that were collected over time.

  66. 13:13

    The Qodo Merge tool, the code review tool collects best practices over time, and this, the context that is coming from the Qodo Aware and the best practice coming for Qodo Merge, helping you write a whatever code generation tool you're using or using directly Qodo tool to make sure that that's in high quality.

  67. 13:30

    But recently, like 20 minutes ago, uh, we released also a, a CLI tool, and that is meant to help you with the workflow parts. The workflow is already integrated into our solution, but this like really, uh, a workflow on, on steroids.

  68. 13:48

    There's another thing very important that is coming out in two months and stay tuned, follow us, whatever, any platform that, that you want. I kind of like, uh, didn't wanna show it right now.

  69. 13:57

    So CLI tool. Why? I hope like I gave enough like hints, um, and, uh, so far, and the main reason is putting AI to work across the SDLC. And I remind you that starting five minutes ago, you can go download it and play with it.

  70. 14:11

    It's a soft launch. Uh, we're gonna announce about it more deeply like in, in, in a, in a couple of weeks. So interestingly, who, who saw Simon's like, uh, talk?

  71. 14:19

    Like he's magnificent. What? Some people didn't? Okay. So please go. Uh, go out of the room, I mean, and, uh, and then come back after and watch onli- uh, uh, online later.

  72. 14:30

    And it's amazing, like when he wanted to demonstrate something, did he open like one of the IDEs? He did not. Like, everything he did was with, with a CLI tool.

  73. 14:40

    Like that's... If you are a developer, like that's idea that you can run it in the background, you can run it in different workflows, you can make workflows. That's part of what I'm, what I'm going to show you.

  74. 14:50

    And, and basically with, with Qodo, uh, the CLI tool, which is live, uh, live here, uh, then I can, for example, chat with it. And very interestingly, that is ver- quite...

  75. 15:02

    Like there are a few tools that we talked about before that you can do that, but interestingly, you can call agents. For example, I can call the, the Explain agent.

  76. 15:09

    It's a dedicated agent. Uh, due to lack of time, I'm not gonna show that. I'm gonna show something else. So how does... How did we, uh, made this Explain agent?

  77. 15:18

    You can actually, with Qodo the agent, you can make other agents, and soon I'm gonna show you how you also create workflows and things like that. So basically, I can go and ask it if the internet works, uh, well enough.

  78. 15:31

    Uh, you can go and ask it, "Please create an agent with this goal." And, and by the way, you can, uh, uh, like give much more information than what I'm doing right now.

  79. 15:40

    I'm asking for it to create a review agent. Okay? If the internet works, it's, it's gonna start working. It's initializing. MCP is gonna search what it can do, what it cannot do, and it...

  80. 15:51

    now it's, it's going to work for, for a few minutes, okay? Like if I would ask you to do it, you would do it, take it in two hours.

  81. 15:56

    So let's give it two minutes. Meanwhile, I'm going to, to progress and, and, uh, and, and tell you that the... I think one of the interesting things, uh, uh, about it, it- it's working.

  82. 16:06

    So one, uh, most agents, like I- I can't s- think of n- of n- of one that won't, would have tools. Uh, for example, the Qodo Cover agent. Uh, by the way, we have an open source for Qodo Cover.

  83. 16:20

    Uh, it's a tool that automatically, uh, generates coverage, coverage. Uh, it i- increases the coverage, automatically increases coverage, but you don't need anymore to go download this open source, et cetera.

  84. 16:32

    You just get it as part of the, uh, uh, was as one of the agent. Okay? So, uh, but if you do wanna play with the open source, et cetera, it's, it's, it's fully open source.

  85. 16:41

    Okay. Back, back to my point. I can go and call Qodo Cover, which uses different tools. For example, it uses tools related to coverage reports. Uh, and, and you...

  86. 16:49

    I can look what I can do in a command line. I can tell it, "Dump all the logs somewhere." Right? How do you do that with an ID plugin?

  87. 16:56

    You can, but y- you need to hack it. Hack it. But the interesting thing, what can we do with command lines? And many here are, are developers. You can pipe, you can pipe, uh, uh, tools together, agents.

  88. 17:07

    Uh, you can pipe like, uh, uh, uh, workflows, et cetera. So I wanna run Qodo Code, generating code. I then wanna run the cover that like, you know, incr- increase the coverage testing.

  89. 17:18

    Maybe the code will generate a few tests, but I have an agent that is specializing in, in increasing coverage. And then I want the review agent that is now cooking and finished.

  90. 17:26

    Uh, um, and I'll, I'll go back to it in a, in a minute if I have time. Now, this is not no agent to agent communication. I did a pipeline.

  91. 17:34

    I just passed the result. But what if we can... And, and, and before I exemplified with MCP, but what if we can invent a command line for A2A? So those agents are, will run in parallel as you go, and they will have some communication, uh, between them.

  92. 17:51

    So I actually have to say that yesterday we had like a 20, uh, 50, uh, like people like, like those that you had, uh, we... The ta- uh, speakers and I ask who is using A2A.

  93. 18:02

    Uh, no one is using it, so it's still a way to go. But this is an exam- Oh, yes. Uh, come talk to me later so I will know one person.

  94. 18:09

    Great. Thank you. So- Uh, uh, and then I'll edit all the posts. Uh, I know one person and, and, uh, and, and... But, but we're getting there. Okay? This is an example.

  95. 18:18

    You need multiple agents. You need A2A. You, you would want to use it if you want discover- discoverability between agent or handshakes. So this is a good example if you have many agents, and you can run them in the background, et cetera.

  96. 18:31

    So in this case, the agent finished, and I have the definition, uh, I have the definition of the, of the agent, uh, over, over here. Uh, it's called Merge Agent, and you can see like w- w- like all...

  97. 18:43

    Everything, the instruction that was made for it, uh, like what tools it is, it is using. Um, I'm gonna add one that, uh, sometime. I wanna be very clear it's there.

  98. 18:52

    And by the way, it also has like a, a output scheme. I'm gonna remove that, move that for a second, okay? And then I can go back, uh, and, and just call it.

  99. 19:03

    It was called Merge Review. Uh, Qodo Merge, uh, Review Agent. It was just generated a minute ago, and it will review. Uh, I didn't give it any... You see that there's a parameter automatically created like, uh, uh, uh, that, that, uh, uh, I can tell it what to review, et cetera.

  100. 19:21

    Give instruction. It's very flexible, but I didn't. Uh, okay. So, uh, I'm almost lack of time, so I won't, uh, show you, but what I, I just will tell you that basically I can tell it to use best practices from the organization, important for vibe coding with confidence, and I can also tell it to have some success

  101. 19:37

    or, or, or failure criteria to know that something happened well or not. For example, if it managed to get some coverage. Okay? That's how you work with, with confidence.

  102. 19:48

    Um, yeah. And you can connect things to pre-commit, post-commit. Again, everything you can do, you can do automatically. Very interestingly, and, and with that I'll almost reach the end.

  103. 19:59

    Very, very interestingly, you can say the CLI tool is very lame.

  104. 20:05

    But notice... Did you notice what, what Simon did? He used the CLI, but eventually checking content he, he used different interfaces. So you can, from the CLI itself, automatically generate interfaces that are fitting for your task, and one of those, for example, is a interface for, for code review.

  105. 20:25

    So who cares anymore about IDEs when you can have flexible IDEs? And with that, I'll finish, and I'll just tell you that I think we're going into a future where we're gonna have swarm of agents, each one specializing and having different credential, different, uh, best, best practices, et cetera.

  106. 20:41

    And, uh, and, and it's coming really fast, 2025, 2026. Thank you. [clapping] [upbeat music]