In today’s issue:
NVIDIA AVO hits 100% on ARC-AGI-3 TrueFoundry open-sources TrueForge harness
Chroma ships Foundation agent memory
Self-improvement gains vanish when reshuffled
Ornith-1.5 self-improves to Opus level dots3-note runs for days
DeepSeek adds vision to V4-Flash
Stripe acquires OpenRouter
Cursor rebuilds Git storage as a database
Harvey post-trains Tenet on Kimi K3
Slack Code makes coding multiplayer
LEGO-RL trains agents in Claude Code
And all the top AI dev news, papers, and tools.
Top Stories #
NVIDIA AVO Solves ARC-AGI-3
NVIDIA’s general-purpose coding agent AVO scored a perfect 100.00 RHAE on the ARC-AGI-3 public set, clearing all 183 levels across all 25 environments with no instructions, rules, or stated goals.
The harness carries the result: Claude Opus 5 alone scores roughly 30% on the same benchmark. Wrapping it in AVO takes it to 100%, which NVIDIA frames as evidence that system design, not model capability alone, unlocks frontier long-horizon performance.Agentic variation loop: AVO cycles through inspect context, plan the next change, implement, evaluate with a scoring function, then diagnose and repair from failed attempts, committing accepted candidates into a growing solution lineage.Persistent memory and a supervisor: Prior implementations, evaluation results, compiler output, and reasoning are retained so the agent resumes from current state, while a separate monitor watches for stagnation and conditionally redirects the main agent.Transfer across domains: The same architecture was originally built for CUDA GPU kernel optimization and moved to interactive reasoning unchanged, using about 12% fewer environment actions than the prior VISTA baseline.
TrueFoundry Open-Sources TrueForge
TrueFoundry open-sourced TrueForge under an MIT license, a vendor-neutral agent harness that acts as the runtime layer turning an LLM into a working agent.
Batteries included: MCP tools, a skills registry, sandboxed execution, human-in-the-loop approval gates for sensitive actions, subagents, durable state for long-running tasks, and step-level tracing all ship in the box.Vendor-neutral by design: It runs on your own infrastructure against any commercial or open-source model, with per-task switching, and the TrueFoundry AI Gateway is optional rather than required.Cost is the pitch: On DevRev’s Enterprise-Bench, evaluated blind, TrueFoundry reports roughly 30% lower cost on identical tasks with the same model, and up to 75% savings when routing to open-source models with accuracy matched.Traction: The repo has cleared 2,800 stars since the August 19 launch. Star it if you want to follow the project.