I Built a Tiny AI Buddy
👉 Agent Buddy on the VS Code Marketplace
When you run multiple AI coding sessions at once, context switching gets expensive.
I kept finding myself asking:
- Which session is actively editing?
- Which one is stuck thinking?
- Did that terminal run finish?
- Is this session healthy or silently failing?
So I built a small VS Code extension: Agent Buddy — an animated sidebar companion that reacts to coding activity in real time.
It’s basically a status dashboard disguised as a character.
And honestly? It makes long, text-heavy vibe coding sessions feel lighter.
Why I built it
Two reasons:
-
Parallel AI sessions need fast visual feedback
- I often have several AI workflows running in parallel.
- Reading logs or scanning text for each session is slow.
- I wanted an at-a-glance signal for “what’s happening now.”
-
Vibe coding can be textually intense
- Constant walls of text are mentally fatiguing.
- A simple visual layer breaks that up.
- It turns monitoring from “reading everything” into “glancing and deciding.”
What Agent Buddy does
Agent Buddy lives in the VS Code Activity Bar and updates based on editor behavior.
It reacts to 9 states:
idlethinkingplanningsearchingeditingterminalsuccesserrorsleeping
Each state has its own color, expression, and animation profile.
So instead of parsing status from text, I can infer state instantly:
- green + energetic = good
- red + alert = fix now
- low activity = waiting or idle
All 9 states in order: idle, thinking, planning, searching, editing, terminal, success, error, sleeping.
How it works (high level)
- Built as a VS Code extension with a custom webview panel.
- Extension listens to workspace/editor signals (edits, diagnostics, terminal, etc.).
- Signals map to a small state machine.
- Webview receives state updates and animates the buddy in SVG/CSS.
- Includes an inactivity timeout that moves to
sleeping.
I also added a screenshot generation flow so docs/marketplace assets stay aligned with actual states.
Unexpected benefit
I originally built this for utility, but it also improved flow.
The UI gives me:
- less cognitive load
- faster context recovery after interruptions
- a little emotional pacing during long coding runs
It sounds small, but in long sessions the difference is real.
If you’re building with AI a lot
If your workflow is mostly text and you’re juggling multiple agents, add one visual signal layer.
It doesn’t have to be a character.
It can be dots, bars, badges, whatever.
The key is this:
Use visuals to compress status, so your brain can spend cycles on decisions instead of parsing.
What I’d add next
- Per-workspace buddy instances
- Session labels for parallel agent runs
- Optional compact mode
- Better metrics around task duration and success/error ratios
Try Agent Buddy
If you want to install it and use it in your own workflow, grab it from the VS Code Marketplace:
👉 Agent Buddy on the VS Code Marketplace
If it helps your flow, I’d love feedback on what state or signal you’d want added next.
If you’re curious, I can share a follow-up post with the exact event-to-state mapping and webview animation approach.
Top comments (2)
The 9-state detection is the right level of granularity — you need enough states to be useful but not so many that you're spending more time interpreting the dashboard than just reading the terminal.
I've been working on the exact same problem but from the terminal side rather than VS Code. My approach uses pattern matching directly on PTY output — detecting braille/asterisk spinners for "thinking", prompt patterns for "waiting for input", error keywords for "error", etc. The advantage is it works with any CLI agent (claude code, gemini cli, codex) without needing IDE integration.
The parallel session monitoring is the real pain point though. Once you have 3+ agents running, the question "which one needs me right now?" becomes the bottleneck. Sounds like we arrived at similar solutions from different angles.
Built mine as a standalone terminal IDE called PATAPIM (patapim.ai) — 9 terminals in a grid, each with state detection. Would be curious to compare detection accuracy approaches.
The detection accuracy in this vscode extension is pretty poor as it is reliant on hooking into vscode’s terminal use notifications, file edit notifications etc rather than the actual agent. Copilot Github Agent doesn’t expose agent events so you can’t get direct state information.
I’ll defn checkout CLI solutions as I’m increasingly finding that I don’t need a code editor because I’m not editing code anymore just vibe coding.