AI Agents - February 2026
It’s been three years since the initial GPT-4 boom, and the hype has finally started to settle into actual utility. We’ve moved past the "AI will replace us" panic and into a phase where the distinction between a simple chatbot and a useful agent has become much clearer.
Agents vs Chatbots
The shift from chatbots to agents isn't just a marketing distinction; it's a change in how we interact with the system. I recently revisited this talk which hits on the core difference.
A chatbot is essentially a passive interface. You prompt, it responds, you iterate manually. An agent is more integrated into the workflow. It's less about "writing a function" and more about "handling a task." We're starting to see systems that can take a high-level goal, look at the codebase, and propose a set of changes that actually make sense in context.
The Feedback Loop
The real progress hasn't just been in the models themselves, but in how they are connected to our development environment.
The "feedback loop" is what makes these systems reliable enough for daily use. Instead of just generating code and hoping it works, agents now live inside the terminal and IDE where they can run unit tests and check CI logs. Being able to self-correct based on compiler errors or test failures—before the human even sees the first draft—removes a lot of the manual "copy-paste-debug" cycle that was so common a couple of years ago.
From Tools to Workflows
We’re also seeing a shift in how we categorize these capabilities. It’s less about a specific "AI tool" and more about "workflows."
When I use an agent now, I’m looking for specific "skills"—like refactoring or migrations. It’s a higher level of abstraction. I don’t need a tool to just "write code"; I need it to understand how to safely move logic across the project without breaking dependencies. It's about orchestrating the boring parts of software engineering so we can focus on the architecture.
The Current Stack
The ecosystem in 2026 has matured quite a bit:
- Gemini CLI & Claude Code: These have become standard for terminal-based work. They're good for quick, multi-file edits where you need the agent to understand the broader project structure.
- Crush: Useful for tackling technical debt. It’s specialized for identifying and cleaning up legacy patterns, which is often more helpful than just generating new code.
- VS Code vs. Antigravity: VS Code is still the standard, but Antigravity has found a niche by focusing entirely on agentic workflows, prioritizing the interaction between the developer and the agent over the traditional text editor UI.
- Jules (jules.google.com): Inside the Google ecosystem, Jules is the go-to. It has deep context on internal libraries and team patterns, which makes it far more effective than a generic model for repo-specific tasks.
Looking Ahead
Our role is becoming less about the syntax and more about verification and constraints. The value is in knowing what to build and how to ensure the agent's output meets the project's standards. It’s a different kind of engineering, but it’s one that rewards a deeper understanding of system design over just knowing a specific language's API.