Redbird Labs
World Champion · IEEE CoG 2026
Redbird's competitive Pokémon AI took first place in the Championship Track of the 2026 VGC AI Competition at the IEEE Conference on Games, finishing 205 Elo ahead of the next-best agent in the world.
First place in the Championship Track of the VGC AI Competition at IEEE CoG 2026, held this year in Madrid.
The VGC AI Competition at the IEEE Conference on Games is the academic benchmark for Pokémon VGC agents, and the Championship Track is its main event: the complete game, where every agent has to build its own team and then battle the field head-to-head, with results distilled into an Elo rating. Loblolly finished on top.
It didn't just finish on top, it dominated. Loblolly's final Elo of 1616 was 205 points clear of the second-place agent at 1411. In Elo terms, a 205-point gap means Loblolly would be expected to beat the next-best agent in the world 76.5% of the time.
The Battle Track is a different kind of test. Every agent pilots randomly generated teams through a single-elimination bracket, where one lost round sends you home, so the organizers ran the bracket 100 times. Loblolly won four of those brackets outright. The Championship Track adds the hardest decision in Pokémon: building the team itself. Each season is 100 games against the field on a randomly generated roster, with Elo updated after every match. The whole season is then replayed on fresh rosters with Elo carrying over, so the final rating reflects how each agent fared across many rosters, not just one that happened to suit it. More choices for the agent and less room for luck: that's the challenge Loblolly won.
Official results as posted on the VGC AI Competition wiki.
| Rank | Agent | Elo | Overall |
|---|---|---|---|
| 🥇 1 | Loblolly | 1616 | 2.56 |
| 🥈 2 | SYbot | 1411 | 1.55 |
| 🥉 3 | DaehoV3 | 1409 | 1.53 |
| 4 | Example 3 | 1166 | 0.33 |
| 5 | YY봇 | 1163 | 0.32 |
| 6 | ParanoidBot6 | 1155 | 0.28 |
| 7 | AlakazAI | 1150 | 0.25 |
| 8 | Jirachi | 1067 | -0.16 |
| 9 | Ariadne | 977 | -0.60 |
| 10 | Choi | 880 | -1.08 |
The Overall score factors in both Elo and a time component: how long each agent spends thinking.
| Rank | Agent | Tournament wins | Match wins | Avg. rank |
|---|---|---|---|---|
| 🥇 1 | SelfPlayFlat2026 | 23 | 187 | 4.05 |
| 🥈 2 | DaehoV3 | 22 | 137 | 5.23 |
| 🥉 3 | Example 3 | 15 | 133 | 5.31 |
| 4 | AlakazAI | 11 | 125 | 5.32 |
| 5 | YY봇 | 11 | 117 | 5.16 |
| 6 | Ariadne | 6 | 89 | 6.38 |
| 7 | Loblolly | 4 | 86 | 5.85 |
| 8 | SYbot | 3 | 88 | 5.94 |
| 9 | Jirachi | 3 | 77 | 6.32 |
| 10 | ParanoidBot6 | 1 | 38 | 7.72 |
| 11 | Choi | 1 | 23 | 8.72 |
| 12 | Example 11 (random agent) | 0 | 0 | 12.00 |
Every agent pilots randomly generated teams through a single-elimination bracket. The bracket was run 100 times, and Loblolly won four of them.
Loblolly is Redbird Software's competitive Pokémon AI. It plays VGC, the official doubles battling format of the Pokémon video games, and it's one of the most difficult games ever handed to a machine: a staggering number of possible teams and moves, hidden information, dice-roll randomness, and opponents who adapt to you in real time. Chess engines see the whole board. Loblolly doesn't get that luxury, and that's exactly why we picked the game.
It's named for the loblolly pine, a North Carolina native, just like our cardinal.
Loblolly won the Championship Track of this year's IEEE CoG VGC AI Competition, and has its sights set on the PokéAgent Challenge's live leaderboard next.
The academic benchmark for VGC-playing agents, held at the IEEE Conference on Games. Entries face off in battling and championship tracks. Loblolly took first place in the Championship Track, where every agent builds its own team.
Started as a NeurIPS 2025 competition and now run as a live leaderboard, the PokéAgent Challenge pits agents from research labs and independent teams against each other in ranked singles Pokémon Showdown matches. Loblolly will be on it the moment a VGC ladder opens, and our singles agent will join when it is released.
Redbird exists to turn messy, high-stakes technical problems into reliable software. A game full of uncertainty, randomness, and adaptive opponents is the same kind of problem we solve for clients, just with more Pikachu.
Most game AI is built for games where you can see everything. Pokémon hides half the board. Loblolly was designed around uncertainty from day one, not retrofitted for it.
Before any change ships, Loblolly plays orders of magnitude more games than a human could in a lifetime. Every improvement is validated with data. No hunches, no hero moves.
The project is led by David Kinlaw, who holds an M.S. in Computer Science with an AI specialization from Georgia Tech and has 13+ years of engineering experience across startups and established tech companies.
It's a fair question, and the answer comes down to format. Loblolly plays VGC, the doubles format used in official Pokémon tournaments, where each side has two Pokémon on the field at once. The PokéAgent Challenge's live leaderboards currently rank singles formats only, so there's no VGC ladder for Loblolly to climb yet.
The PokéAgent server already lists a VGC ladder, currently marked “Coming Soon” with no battles played yet. When it opens, Loblolly will be there.
Our singles agent is in development. Singles is a genuinely different game. With one Pokémon per side instead of two, team building, move choice, and even what counts as a strong position all change, so a doubles champion can't simply be dropped onto a singles ladder and expected to win. Instead, we're adapting Loblolly's engine into a dedicated singles agent, and the goal is the same one that won us a world championship: build the strongest agent in the world, then let the ratings prove it.
The interesting part, how Loblolly actually decides what to do, stays in-house until the season is over. Its architecture, training methods, and everything in between are trade secrets for now.
That said, we believe in giving back to the community that makes these competitions possible. Once the results are in, we may be open to presenting our methods at future competition conferences.
If you're a competitor reading this: good luck out there. If you're anyone else: we'd love to talk about what this kind of AI could do for your business.
The same engineering that won Loblolly a world championship (decision-making under uncertainty, rigorous evaluation, custom AI built to win) is what Redbird brings to client work.
Get in touch