A developer built soksolv, a Sokoban solver, entirely through a single long Claude Code session plus roughly 430 subagents, without writing any code by hand. It outperforms every published solver on both standard benchmark suites, including 24 world-first automated solves, and runs about 3.2x faster in total time at an API-equivalent cost of ~$3,500 and 15,000 vCPU-hours.
Subagents
子代理
Parallel task-execution units spawned and orchestrated by a main agent to divide large engineering work.
Curated Intelligence
The SenseNova team introduced the RoboRSI framework, which significantly enhances embodied AI performance by employing multi-point end-effector prediction, planning/feedback subagents, and a recursive self-improvement loop while keeping GPT-6 Astra weights frozen. This approach achieved an average score of 56.83 on RoboDojo (vs. baseline 28.97) with a 50.83% task success rate, demonstrating that superior agent system design can outperform mere model scaling.
Anthropic released Claude Haiku 5.5, a small model optimized for high-volume tasks with costs reduced by ~75% compared to Haiku 4.5, serving as a coding subagent for larger models. Additionally, Sonnet 5.5's cache read prices were halved, lowering the cost of most agentic workloads by ~20%.
A developer used Opus 5.5 to autonomously clone SimCity 2000 into a WebGL low-poly version within days using a single prompt. The model leveraged Ghidra for decompiling original PowerPC code, deployed 27 subagents for parallel execution, and built a Python verification harness running 10,000 game days to ensure identical state logic. This case demonstrates the practical capability of frontier LLMs in complex reverse engineering, multi-agent collaboration, and self-verification loops.
OpenAI has released a major update to its Agents SDK, introducing a model-native harness and built-in secure sandboxing to address production deployment challenges. The update synchronizes core features like code mode and subagents across Python and TypeScript, while expanding support for third-party sandbox providers.
OpenAI officially released GPT-5.4 mini and nano, targeting high-volume workloads with optimal cost-performance ratios. GPT-5.4 mini significantly outperforms its predecessor in coding, reasoning, and multimodal tasks while running over 2x faster, approaching flagship GPT-5.4 performance on benchmarks like SWE-Bench Pro. The nano variant is optimized for low-latency, low-cost tasks such as classification and simple subagent operations.