AgentGarten: code-defined simulated worlds where agents learn from acting
mhdfaran · x · 2026-10-10
MirroS introduces AgentGarten, a research project built on the idea that agents need to learn from taking actions, not just answering questions.
- Executable worlds: environments are defined in code — code decides how the world changes, making objects, rules and world states directly queryable and editable, unlike video-generation-based world models.
- Real-time neural renderer: a neural renderer fills the visual gap of procedural scenes, letting agents act on what they see and write down what they learn.
- Hide-and-seek demo: agents evolved strategies on their own — building shelters by round 4 and using ramps by round 10.
The project aims to provide an open-ended environment where agents can turn action into experience across navigation, manipulation and tool use.
More from coding & agent
- Open-source typesafe-computer-use cuts agent cost 155x by skipping screenshots and frontier models — thisguyknowsai · 2026-10-10
- Only 11% of 100 Assessed AI Agents Pass Security Baseline; 98% Show the Lethal Trifecta — bibryam · 2026-10-10
- A 12-stage roadmap to becoming an AI automation engineer in 6 months — thisdudelikesAI · 2026-10-10
- Dev asks Claude Code for multi-account support, citing Codex as example — RileyRalmuto · 2026-10-10
- Indie dev open-sources AI design tool: natural-language editing, multi-platform templates, Obsidian plugin — vista8 · 2026-10-10
- This Linear-style video Skill turns a product link into a cinematic 30s launch film — yihui_indie · 2026-10-10