Podcast Summary: BONUS – Building Reliable Software with Unreliable AI Tools (with Lada Kessler)

Podcast: Scrum Master Toolbox Podcast
Host: Vasco Duarte
Guest: Lada Kessler
Date: November 28, 2025

Episode Overview

In this bonus episode, Vasco Duarte talks to software developer Lada Kessler about the challenges and breakthrough moments in building reliable software with AI-assisted coding tools—even when those tools are inherently "unreliable." Lada shares stories from the frontline of AI-driven development, discussing practical strategies, anti-patterns, and the importance of developer judgment in the age of stochastic, unpredictable LLMs. The conversation explores the evolution of coding methods, from "vibe coding" to intentional, agentic swarm programming, and offers concrete advice and resources for listeners experimenting with AI in their development workflows.

Key Discussion Points & Insights

1. The Concept of “Vibe Coding” vs. Traditional Coding

[02:25–04:08]

• "Vibe coding" refers to coding without deeply understanding or checking the output—essentially letting the AI generate the code while the developer pays little attention.
• Lada highlights the risks: “Please, please don’t do that. Bad idea.” (B, 03:31)
• Catastrophic tales, like accidentally deleting core business databases, show the dangers of unchecked AI-generated code.

2. A Personal Turning Point: Becoming Open to AI Assistance

[05:13–10:18]

• Lada was once skeptical of using AI, finding earlier browser-based tools unreliable and unhelpful.
• The real breakthrough occurred when she used an IntelliJ DevOps plugin to encapsulate a project’s context and provide the entire codebase to the AI. This enabled her to build a working Angular frontend—a tech stack she was previously unfamiliar with:

  "I built a whole front end port for myself this way, and it worked. … I got compliments from my colleagues for my design skills." (B, 08:56)

• The experience revealed how AI can empower developers to gain new skills rapidly if they already possess strong engineering foundations.

3. From Siloed to Agentic (On-computer) Development

[10:18–11:34]

• Tools have evolved from browser-bound code generation to agentic coding tools that sit within the developer’s environment (e.g., Cloud Code, Codex, Cursor, Windsurf).

  “Now you bring it, agentic brings it to your computer and now you can use your tools and you’re so much powerful because of it.” (B, 10:18)

• This means the AI can directly read, edit, and write files—greatly enhancing productivity for experienced developers.

4. Learning to “Prompt” and Formulate Problems

[14:13–17:08]

• Lada describes a pivotal lesson: Avoiding “answer injection” or unwittingly limiting the AI by the way you ask questions.

  “You’re limiting yourself without knowing you’re limiting yourself.” (B, 15:37)

• When Lada reformulated requests and described the environment more precisely, the AI surfaced unfamiliar yet powerful tools (e.g., AppleScript, tmux).

5. Swarm Programming: Multiple AIs Collaborating

[18:30–24:39]

• Lada experimented with “swarm programming,” running several AI instances (e.g., architect, developer, tester) in parallel, enabling them to communicate via scripts and role-based tasking.

  “I have a command that I can give them very quickly to say, ‘Hey, we’re a team of human and cloud codes. We can see who’s around.’ … I was like, 'Oh my God, this is a swarm.'" (B, 19:45)

• She notes that true agentic or swarm development requires additional orchestration, as simple role-based swarms often need more nuanced guidance.

6. The Reliability and Limitations of AI Coding

[24:39–31:10]

• Never trust the output of a single LLM. Lada cross-validates output from one AI with another for critical tasks, leveraging each to review and refine the other's work.
• Emerging frameworks like AIDev, VMAD, and SPEC Kit attempt to bring order and repeatability to AI coding, but reliability challenges remain.
• Lada contemplates if AI can ever truly manage complexity or if skilled humans will always need to guide architectural and refactoring decisions:

  "Right now, the biggest problem... is if I just let it do things, it’ll just run itself to the wall at creative speeds because it’s really good at running." (B, 28:39)

7. Pragmatic Use: Adjusting Speed and Rigor Based on Context

[31:10–33:03]

• Lada advocates for situational awareness—going “fast” with AI when safe (e.g., prototypes, hackathons), but exercising experience-driven judgment and discipline for production code.

  “You need to... be attuned to environment. Like, you can go fast and slow, and sometimes going slow is bad because if you’re on the highway, you’re going to get hurt.” (B, 31:45)

• Example from the Hack for Good event: Taking risks with AI and rapid iteration enabled high achievement within tight deadlines, as long as refactoring and oversight were available.

8. Resources & Community Learning

[33:27–36:45]

• Lada created a talk for Calgary Software Crafters cataloging patterns, anti-patterns, and obstacles for AI-assisted coding. The talk and interactive website are especially valuable for those moving from beginner to advanced AI-assisted development.

  “The whole aim of this talk is kind of give you the foundation and specific skills... I kind of distilled my experience in a year into that talk.” (B, 33:40)

• Lada will present a live workshop at DEFCON New York (November 18) and shares her contact details for further engagement:

  “LinkedIn would probably be the best resource. So I’d reach out there.” (B, 35:32)

• Links and resources will be in the show notes.

Notable Quotes & Memorable Moments

• On “Vibe Coding”:

  "Vibe coding is basically the extreme ... you just talk to a computer, you say do this, do that. You don’t really care about code. ... Please, please don't do that. Bad idea." (B, 03:31)

• On Gaining New Skills via AI:

  “Oh my God, I don’t need to—you kind of got a new skill for free, right? Suddenly it was there.” (A/B, 09:13–09:18)

• On Swarm Programming:

  "I could have any AI with [my] machine talk to each other and work together if I want to. And I've been playing with it a little bit. ... I was like, oh my God, this is a swarm." (B, 18:24–20:16)

• On Not Trusting a Single LLM:

  “Never trust the output of a single LLM ... I do this often with two clause or with codecs lately and have them give feedback to each other and self correct.” (B, 22:23)

• On Managing Complexity:

  “Can it manage complexity? ... If I just let it do things, it’ll just run itself to the wall at creative speeds because it’s really good at running. So I have to be there managing complexity for it.” (B, 28:39)

• On the Experimental Frontier:

  "We're in the era of gentlemen scientists. ... Newbies can make big discoveries right now because like nobody, nobody knows what it is." (B, 26:38)

Timestamps for Important Segments

• [02:25] – Vibe coding explained; the danger of unchecked AI-generated code
• [05:18] – Lada’s critical moment—using AI tooling effectively on the job
• [10:18] – The agentic evolution of code tools: from browser-based to local environments
• [14:13] – Overcoming self-imposed limitations by learning to prompt differently
• [18:30] – Swarm programming: orchestrating multiple AIs on a single machine
• [24:39] – Reliability, anti-patterns, and the wisdom of double-checking AI outputs
• [31:10] – Situational speed and rigor; learning when to go “fast and loose” or careful
• [33:27] – Talk, interactive resources, and community engagement links

Tone & Language

The conversation is lively, candid, and full of friendly curiosity. Both Lada and Vasco balance realism about the limits and failures of AI tools with excitement and encouragement to experiment and share findings—echoing a community-oriented, growth mindset.

Further Resources

• Lada’s Calgary Software Crafters talk (video and interactive patterns website) – See show notes
• DEFCON New York Workshop on AI-assisted coding – November 18, 2025
• Lada’s LinkedIn – recommended for networking and updates

Closing Thoughts

Lada and Vasco jointly emphasize that we are at the beginning of a new era for software development—a period that rewards curiosity, experimentation, and openness to share discoveries. As AI-assisted coding evolves, the developer’s craft will shift from solo creator to orchestrator and judge, ensuring reliability amid new forms of speed and collaboration. Listeners are encouraged to experiment, connect with the community, and contribute to this ongoing story.