TovanaEngine: a local world model trained on 50K+ real SWE-bench coding-agent runs
Decent-Ad9950 · reddit · 2026-09-20
A developer open-sourced TovanaEngine, a local "world model" that gives coding agents experiential memory.
- Trained on 50K+ real coding-agent runs from SWE-bench Verified across 25 models, plus behavioral signals (merge rates, review time, repo size), yielding over 1M step-to-step transitions
- Rationale: a coding agent only sees its current run and lacks knowledge of how similar runs usually unfold; the model helps it make better decisions and waste less reasoning on bad paths
- Runs locally in Docker, no account, nothing leaves your machine; available via MCP and HTTP API, with full details in the repo
Author calls it an early-stage project with more models on the way.
More from coding & agent
- Google Rolls Out Browser Use to All Antigravity Users, Raising Doubts About Gemini Desktop App — testingcatalog · 2026-11-12
- OpenRouter launches Extractions API: typed JSON field extraction from text, images and PDFs — michellechen · 2026-11-12
- AgenticRL-Bench debuts: best agents pass only 14% of real-world agentic RL tasks — cuijiaxun · 2026-11-12
- SemiAnalysis: Persistent agents now eat 45% of our AI spend, and Perplexity's harness beats Claude Tags — dylan522p · 2026-11-12
- Gemini Agent Spans Workspace, Mobile, Desktop, and Slack via @-Mention — samcharrington · 2026-11-12
- Carbon + /swarm in Antigravity: Big Model Thinks, Lighter Models Work — samcharrington · 2026-11-12