choong pw

Over the past year, AI‑assisted development has grown beyond autocomplete suggestions and chat‑based help. We’re now entering what I see as the agentic phase of software engineering — where coding assistants behave more like collaborators. Instead of responding to one‑off prompts, they internalize context, follow defined rules, and make decisions on your behalf.

This shift didn’t hit me immediately. It became clearer as I experimented with new frameworks and tools, especially during the hackathons I joined recently.

The discovery of steering docs

While building apps on Atlassian Forge, I stumbled upon something interesting: when running the forge create command, it generated a file called AGENTS.md. I know what this file does, but, wasn’t expecting it to come out of the box during project initialization. It caught me off guard, and I didn’t even bother to read it.

Steering docs (in Kiro) respect AGENTS.md act as rules of engagement for AI. They describe how an agent should operate, what to reference, what styles to follow, and which guardrails to respect. For the first time, I didn’t have to craft elaborate prompts, and just let the tools to do it’s thing. Guess what? it works as expected. More later.

Now design intent is becoming codified. Now, instead of programming behavior through code alone, we’re now programming behavior through context.

MCP for self‑healing code

Atlassian didn’t stop at AGENTS.md — they also introduced an MCP called forge‑knowledge. This component lets the AI reference Forge’s documentation internally. The result is fascinating: when something breaks, the AI doesn’t just surface errors — it often self‑heal, pulling in the right patterns from its own context to fix the issue automatically.

It often self‑heal, pulling in the right patterns from its own context to fix the issue automatically.

That’s a huge step forward from static coding. Instead of “assistive” tools, we’re now looking at adaptive ones — agents that reason with context, cross‑check against documentation, and make informed corrections.

Bridging the gaps with Jedi MCP

As impressive as these setups are, they still have limits. Tools like context7 aim to close the gap in context awareness across frameworks, but they can’t cover everything. Inevitably, there are always libraries or framework that are missing from the shared context list. Especially the new one.

That’s where the idea for Jedi MCP came from. I wanted a simple CLI tool to extract knowledge, build it into a local MCP, and make it runnable. The goal isn’t to replace larger context managers but to bridge them, empowering more developers to compose their own coding agents instead of waiting for built‑in support.

Recently, my company held an internal hackathon using Microsoft Agent Framework, which still lacks an MCP server. That felt like the perfect chance to test Jedi MCP in a live environment. To my surprise, it worked right out of the box. Watching the agent reference external knowledge and resolve missing context autonomously was one of those rare “wow” moments where automating technical work suddenly feel alive.

Designing for agentic development

What’s clear from all this is that the way we build is changing. Agentic coding isn’t about asking AI for snippets anymore — it’s about designing systems that can think along with you.

You define the world they know through documentation and APIs. You steer their behavior through metadata and intent. And when that foundation is solid, the resulting collaboration feels like working with a teammate who reads the entire spec before writing a line of code. The perfect programmer 😀

This autonomy doesn’t remove the need for humans — it amplifies our ability to design. The better we plan and steer these agents, the more precise their output becomes. That’s why spec‑driven development, which I wrote about earlier, pairs so naturally with agentic frameworks.

Looking Forward

I plan to keep Jedi MCP public, and I hope more developers experiment with building with MCPs. Every hackathon teaches me how much faster and cleaner AI‑assisted engineering is becoming — not just in coding speed, but in clarity of intent.

We’re still early in this shift, but the trajectory is ever changing. Things may change 1 year from now. Nevertheless, it definitely will be for the better if we continue to update ourselve.