What if most of the AI tools promising to revolutionize DevOps are doomed to fail because they only see a fraction of the picture? In this first episode of Humans in the Loop, Andrey and Fernando walk through a timeline of AI tooling from ChatGPT's launch in late 2022 through the breakthrough year of 2025, arguing that context is king and that tools without full access to source code, logs, metrics, and documentation will inevitably hallucinate their way to irrelevance.
Key Topics
From Chatbot to Agent: Defining the Shift
Fernando offers the simplest framing of agentic AI: “an agent can do something.” Unlike ChatGPT or Claude in a browser — where you type a question and get text back — an agent can take actions: run commands, modify files, call APIs, and interact with real systems. This distinction matters because it explains why many infrastructure engineers initially dismissed AI. A chatbot that generates plausible-sounding Terraform snippets without knowing what’s actually deployed is, at best, a faster Stack Overflow. The hosts argue that the transition from text-in-text-out to action-capable agents is what makes AI genuinely useful for DevOps work.
2022–2023: The Autocomplete Era
ChatGPT’s release in November 2022 kicked off the current wave, but for infrastructure practitioners, it was mostly a novelty. By 2023, AI-assisted coding tools started gaining traction. GitHub Copilot gave developers their first taste of AI in the IDE. Fernando recalls an early “aha moment” when Copilot generated test data that included a zip code matching a fake address, demonstrating it was more than simple autocomplete. Cursor also arrived on the scene as one of the first “AI-native” IDEs, purpose-built for AI integration rather than bolting it on as an afterthought. Still, the overall feeling in 2023 was “clever autocomplete” — useful for writing code faster, but not transformative for the troubleshooting and operational work that consumes most of a DevOps engineer’s time.
2024: The Quiet Year of Foundational Work
Twenty twenty-four was surprisingly quiet on the surface. Windsurf appeared, but few other tools made it onto the hosts’ radar. Behind the scenes, however, critical groundwork was being laid. Andrey notes that Anthropic introduced the Model Context Protocol (MCP) in late 2024 — citing a November 25, 2024 date while cautioning that he got it from a Google AI Overview “which you probably shouldn’t” believe. The hosts also note that both Anthropic and OpenAI released desktop applications during this period, which was a meaningful step: browser-based chatbots are sandboxed by design, but desktop apps running locally can access files, credentials, and system tools. For DevOps engineers specifically, Andrey notes that 2024 “hadn’t brought much of the changes” — chatbots still lacked the context needed to troubleshoot real infrastructure problems, and AWS’s early attempts to embed AI into its web console “wasn’t working very well at all.”
2025: The Year Everything Arrived
Twenty twenty-five was what the hosts call a “breakthrough year” with, in Andrey’s words, “an explosion of things” that became practically useful. Fernando notes that Claude Code emerged in early 2025 as a CLI-based agentic coding tool. Andrey initially overlooked it but acknowledges it became much more than “another version of Cursor” once people started using it seriously. Andrey describes Amazon Q CLI arriving in March 2025, and — deeply embedded in the AWS Community Builders ecosystem — he spent significant time with it. The key capability: because it runs in your terminal and inherits your AWS credentials, it can call AWS CLI on your behalf, fetch and process logs, generate infrastructure reports, and even make changes with your permission. This was “actually quite a big deal” for troubleshooting. Andrey notes that Amazon Q Developer CLI has since been rebranded as Kiro CLI. Other notable tools the hosts mention include Cline (an open-source agentic tool that Andrey heard also had spec-driven development capabilities) and OpenCode (an open-source terminal-based agentic coding tool that Andrey describes as similar to Claude Code). Kiro, AWS’s spec-driven IDE, entered preview in July 2025, according to Andrey.
Kiro and Spec-Driven Development vs. Vibe Coding
Andrey spent extensive time with Kiro and sees spec-driven development as the future of AI-assisted coding. Rather than throwing a vague prompt at an AI and hoping for the best — what he compares to Google’s “I’m Feeling Lucky” button — Kiro’s workflow has developers describe a feature, then generates a requirements list, design spec, and implementation specification before writing any code. Combined with heavy emphasis on test-driven development (including integration tests), the output quality was “much better comparing to your normal vibe coding.” Fernando tried to replicate this approach in Cursor by always creating a plan before writing code, which helped but lacked Kiro’s structured enforcement. Both hosts agree that planning mode — now available in Claude Code, Cursor, and other tools — is where developers should start every time.
Context Is King: Why Most AI Plays Will Fail
Andrey offers a pointed prediction: most SaaS tools attempting an “AI play” will fail because they only have a piece of the picture. A monitoring tool like Better Uptime receives alerts but has no access to source code. Even with logs, it lacks the broader context of infrastructure configuration, Slack conversations, and documentation. “In order for that to work, you really need to have everything,” Andrey argues — source code, logs, metrics, communication history, and documentation. Fernando draws a parallel to human engineers: joining a team cold and debugging a crash without context forces you to guess, just like an AI without proper context will hallucinate. The discussion mirrors the problem space B.O.R.I.S is intended to address: giving AI agents the broader operational context needed to be genuinely useful.
Proactive Agents and Real-World DevOps Use Cases
The hosts highlight the emergence of proactive AI agents in early 2025 as a significant milestone — tools that don’t just respond to queries but can monitor systems, detect issues, and reach out with findings. They paint a vivid use case: an agent that watches logs, spots an error, researches the likely cause, and sends a message with its investigation before a human even notices the problem. Andrey gives a concrete example with AWS Health notifications — notoriously dense alerts that might tell you a Lambda runtime is being deprecated somewhere across a multi-account environment, leaving you to manually track down affected resources. An ideal agent would receive the notification, search across accounts, identify affected resources, and present a clear summary with recommended actions. “This is toil,” Andrey says — not value-adding work, just the maintenance tax of running on the cloud. And this is precisely where current AI agents can deliver real value, even if they’re not yet ready to run entire infrastructure autonomously.
What the Hosts Bring: Practitioners Building the Thing
Andrey and Fernando emphasize their unique vantage point: they’re not just using AI tools as consultants but actively building B.O.R.I.S, an agentic AI DevOps teammate that, as Fernando states in the episode, already has clients using it in production. This dual perspective — consuming and creating AI tools — informs their discussions. In upcoming episodes, they plan to cover topics like context management and the practical realities of connecting agents to live AWS environments. While they frequently reference AWS due to their professional background, they note that the concepts are cloud-agnostic and applicable regardless of provider.
Resources
- Model Context Protocol (MCP) — Anthropic — Anthropic’s open standard for connecting AI models to external data sources and tools, discussed by the hosts as a key development from late 2024.
- Claude Code — Anthropic — Anthropic’s agentic coding tool that runs in the terminal, reads codebases, and executes tasks through natural language commands. Discussed as a key 2025 tool.
- Amazon Q Developer CLI — AWS’s agentic CLI tool that inherits local AWS credentials to troubleshoot, query resources, and process logs directly from the terminal. Now succeeded by Kiro CLI.
- Kiro — Spec-Driven IDE — AWS’s IDE focused on spec-driven development, generating requirements, design specs, and implementation plans before writing code. Andrey highlights its emphasis on test-driven development.
- Cline — AI Coding Agent for VS Code — Open-source autonomous coding agent for VS Code with plan/act workflows and MCP-based integrations. Mentioned by the hosts as a notable agentic tool.
- OpenCode — Open Source AI Coding Agent — Open-source terminal-based agentic coding tool supporting multiple AI providers. Mentioned by Andrey as similar to Claude Code.
- GitHub Copilot — GitHub’s AI pair programmer, one of the earliest widely adopted AI coding assistants. Discussed by Fernando as a formative early experience with AI-assisted development.
- Cursor — AI Code Editor — One of the first AI-native IDEs, discussed by the hosts as a tool that helped set the tone for AI-integrated development workflows.