Joel Z
@TheCodeOfJoel
Followers
2K
Following
4K
Media
265
Statuses
1K
I code things sometimes. 🪐 PFP: @Elmi39Project Currently working on https://t.co/RkpnWAzZ7z
Stargazing for our Astrogirl
Joined September 2021
There's still more planned for this project. Follow me to see what's in store! https://t.co/xo4B5Z5Tc5
https://t.co/XK4LnUyfZn
twitch.tv
Gemini Plays Pokémon Crystal | !askgem [!faq !harness] !badges
What a finish! Gemini 2.5 Pro just completed Pokémon Blue!  Special thanks to @TheCodeOfJoel for creating and running the livestream, and to everyone who cheered Gem on along the way.
10
5
65
Announcing: New @Steam Hardware, coming in 2026: Steam Controller Steam Machine Steam Frame Watch our jazzy announcement video and wishlist now: https://t.co/TUKoZdzn9B
3K
18K
119K
Imagine posting that Gemini made zero progress when a) you used 2.0 Pro Exp which came out in Feb and b) it explicitly says it lacks access to real-time info This is w/ 2.5 Pro btw, it takes corrections gracefully & 2nd try is w/ Grounding enabled And 3.0 is around the corner.
lmao Gemini made zero progress on recognizing that 2025 is a real year
0
0
1
Gem is back! ⚡️ Watch Gemini 2.5 Pro take on Pokémon Crystal in a brand-new adventure. How far can Gem go with true agentic freedom? Find out live on Twitch 🎮 https://t.co/aPzOMs3lym
0
0
2
The Making of Gemini Plays Pokemon - @TheCodeOfJoel LLM agent architecture, prompting strategies, and more scaffolding tips for embodied agents in games 📺
0
2
6
Have you ever wondered why we spend so much time making agents play Pokemon? New talks from the PokeAgent Challenge (NeurIPS 2025): @TBFUnreality: Pokemon as an AI Problem (hint: state space complexity and partial info) @TheCodeOfJoel: The Making of Gemini Plays Pokemon
1
5
29
@zeddotdev I've put a ton of effort into making Sumika & Utsuwa a polished experience out of the box, but there's always room to make it better. If you find a bug or have an idea for a feature, contributions are welcome! Looking forward to seeing your PRs.
0
0
0
@zeddotdev Sumika provides a robust HTTP API, making it easy to build custom clients on any platform. While Utsuwa is our reference web UI, you could build a Telegram bot to interact with a local agent, a native app, or integrate ACP agents anywhere. #api #ai #llm #acp #opensource
1
0
0
Try Sumika! 🚀 A stateful API layer for agents using the @zeddotdev's ACP. Adds multi-session management, history & branching. Comes with Utsuwa, a polished reference web client. 🐙 Sumika (Backend): https://t.co/Z5KUrsKqYc 🥣 Utsuwa (Frontend):
github.com
Utsuwa (器), from the Japanese word for "vessel," is the official web client for the Sumika API server. It provides a polished, intuitive user interface for managing workspaces and...
Update: two pieces on the ACP spec by @zeddotdev, open sourcing soon. 1) Stateful workspace-centric API wrapper for any ACP server. 2) First-party web client. Aiming to support any ACP server implementation. Third-party clients are easy thanks to an OpenAPI spec! Screenshots 👇
1
0
0
Claude Code does it natively. Now every agent can. Background Process MCP lets any MCP-capable agent start, stop, and monitor long-running CLI tasks, with full visibility in the built-in TUI. As easy as `npx @waylaidwanderer/background-process-mcp`. https://t.co/Hg1o30JQo3
#MCP
0
0
2
Update: two pieces on the ACP spec by @zeddotdev, open sourcing soon. 1) Stateful workspace-centric API wrapper for any ACP server. 2) First-party web client. Aiming to support any ACP server implementation. Third-party clients are easy thanks to an OpenAPI spec! Screenshots 👇
Just built a prototype headless API for the Gemini CLI! Inspired by Zed's latest update, it was a fun dive into the Agent Client Protocol (ACP) from @zeddotdev. Planning to open-source this, so follow if interested. This unlocks so many possibilities for vibe coding on-the-go...
0
1
3
Tried Comet. It's really special. I tried to switch, but Firefox keeps me locked in: cross-browser tab syncing (mobile too), tree-style tabs, and proper adblock. When Comet matches that, I'm in.
Today we’re announcing a partnership to bring 1Password to Comet, for built-in personal security without interruption.
0
0
3
I want Gemini merch too
Thank you @OfficialLoganK for the Swag! Going to wear it at the live Roo Code weekly podcast tomorrow 9am MT!
1
0
3
Good scaffolding is everything.
Here's how I (almost) got the high scores in ARC-AGI-1 and 2 (the honor goes to @jerber888) while keeping the cost low. To put things into perspective: o3-preview scored 75.7% on ARC-AGI-1 last year while spending $200/task on low setting. My approach scores 77.1% while spending
0
0
2
@TBFUnreality @jakegrigsby @TheCodeOfJoel discusses "The Making of Gemini Plays Pokémon" Joel will share insights about frontier LLMs that interact with complex game environments. His talk preview:
I wrote up the making-of for Gemini Plays Pokémon: how I designed the scaffold so Gemini 2.5 Pro could handle a long-horizon game, what failed, and the lessons that made it work. Full post:
1
1
4
🔴 Final speaker lineup confirmed - PokéAgent Challenge Hackathon starts in 48 hours! NeurIPS 2025 competition featuring two tracks advancing AI decision-making through Pokémon: 🥊 Competitive battling 🏃 RPG speedrunning Research talks Saturday 12-1:30 PM EDT $2k in GCP
4
15
37
Just built a prototype headless API for the Gemini CLI! Inspired by Zed's latest update, it was a fun dive into the Agent Client Protocol (ACP) from @zeddotdev. Planning to open-source this, so follow if interested. This unlocks so many possibilities for vibe coding on-the-go...
2
2
31
Gemini Plays Pokemon just beat Pokemon Yellow Legacy on Hard Mode!!!! - no items in battle - set mode (no free switches) - level cap (no overleveling!!) The Elite 4 went down (no items btwn battles!) after Gemini self-improved its harness (sub-agents) and notes over many runs.
2
4
27
I've been seeing a few posts lately comparing GPT-5 vs o3 on Pokemon that have some misconceptions - to exemplify the problem with these comparisons, one "step" for GPT-5 can include a lookup to game knowledge that instantly solves certain complex puzzles like Cinnabar Mansion
@andrew_n_carr hey heads up, this is inaccurate on multiple levels -- Gemini 2.5 Pro finished in 35k actions (see https://t.co/xMTMQyPpqA Fig 15a), Claude has not finished at all, the definition of a "step" is different across all three (see the 2.5 report), and tools/harnesses are different
3
5
61