Tag
Matryoshka Agent is a hierarchical agent framework that decomposes long-horizon machine learning engineering tasks into a high-level Orchestrator and low-level Sub-Agents, enabling efficient exploration and iterative refinement. It significantly improves performance on complex MLE tasks, allowing a 4B model to match the orchestration of o4-mini and yielding up to 36.7% relative gain on a 30B coder model.
Describes a workflow where OpenClaw delegates complex coding tasks to Claude Code and Codex as asynchronous sub-agents using a local broker script, with results aggregated; the solution is packaged as the open-source Relay-agent repo.
OpenAI unveils GPT-5.6 Sol, a flagship model for long-running autonomous work across applications and enterprise data, featuring Ultra mode with sub-agents for faster, stronger results. The model was used internally to help train Luna and demonstrates significant cost and performance improvements over previous versions.
Prometheus is a local-first AI assistant that uses an orchestrator model via Ollama to delegate tasks to specialized sub-agents, controlling browser, desktop, email, and more. The project is experimental and comes with a warning about security risks.
Claude Fable 5 has been released, enabling autonomous operation with sub-agents, hooks, and persistent memory, building on foundations demonstrated a year ago with Claude Code.
A guide on building a secure agentic system with sandboxing, parallel sub-agents, tool calling with control policies, inference routing, and protection against injection and role escalation attacks, to be published by Evangelos Pappas.
A practical guide arguing that mastering sub-agents requires building four specific workflows in a weekend, covering decomposition, context packaging, verification, and cost control, rather than spending 200 hours on tutorials.
Orkas is an open-source, local-first desktop agent app where a lead agent coordinates specialized sub-agents, each with its own context boundary, using user-provided API keys from various LLM providers.
A developer shares a forked sub-agent repository for pi coding agent that works with a single local LLM slot and limited VRAM, using llama.cpp server and quantized models. The post also discusses performance with the Apex Qwen variant using MTP.
Moonshot AI founder Yang Zhilin released a 40-minute video detailing the training process of the Kimi K2 model, which cost only $4.6 million. In an 8-model real-time programming competition, Kimi K2 took first place, defeating GPT-5.5 and others, demonstrating how a small team can overturn the traditional compute-stacking paradigm through architecture optimization.
This article introduces a framework for using seven specific Claude sub-agents to automate roles such as research, editing, project management, and financial analysis, effectively replacing a high-cost team.
RAO (Recursive Agent Optimization) is an end-to-end reinforcement learning approach for training LLM agents to spawn, delegate to, and coordinate with recursive copies of themselves, turning recursive inference into a learned capability.
OpenAI merges Codex with ChatGPT, introducing GPT-5.6 Sol, Ultra mode, sub-agents, Sites one-click deployment, task management, mobile support, and inline PR review for developers.