Directory

**ARIS** is a methodology-first, Markdown-driven skill system for autonomous ML research workflows. It orchestrates **cross-model collaboration** — Claude Code executes research while an external LLM (Codex, Gemini, or other) reviews work as an adversarial critic. The entire system is files + plain Markdown skills (no database, no framework), making it portable across Claude Code, Cursor, Trae, Codex CLI, and other agents.

AutoHarness tackles a critical LLM agent failure mode: **agents making illegal/invalid actions**.

**47 unique URLs extracted** across 6 categories from 6 HN threads (1,100+ combined points, 418 comments). The HN multi-agent community is skeptical of framework proliferation but hungry for:

The AI agent orchestration market has exploded from $5.25B (2024) to $7.84B (2025), projected to reach $52.62B by 2030 (46% CAGR). The landscape is consolidating around 4 tiers: hyperscaler frameworks (Google ADK, Microsoft Agent Framework, OpenAI Agents SDK, AWS Strands/AgentCore), open-source orchestrators (LangGraph, CrewAI, Agno, PydanticAI, Mastra), protocol standards (MCP, A2A, Agent Skills), and specialized/research frameworks. >40% of agentic AI projects risk cancellation by 2027 due to cost/complexity — the gap between experimentation and production is the central market opportunity.

A mature Ruby multi-agent orchestration framework (~49.3K LOC across 259 non-test source files, 4 gems: SwarmSDK, SwarmCLI, SwarmMemory, ClaudeSwarm-legacy) with sophisticated plugin architecture, 6-pass agent initialization, lazy delegation, Fiber-based circular dependency detection, comprehensive hooks system (13 events, 6 result actions), composable swarms, persistent memory with semantic search, context compaction, and state snapshot/restore. The most architecturally complete open-source agent framework analyzed to date.

LLM agents are amnesiac: they repeat the same failures, miss reusable successful strategies, and cannot automatically apply lessons from past executions. Existing approaches are inadequate:

The paper delivers an unexpected but well-supported finding: **function-calling agents should think briefly, not deeply.** The optimal CoT budget for tool selection is 8–16 tokens — approximately one sentence identifying the function and key arguments. Beyond that, reasoning quality degrades through a documented "dual failure" mechanism where extended thinking causes both function hallucination (the model generates names outside the candidate set) and wrong-function selection (the model talks itself out of the correct choice).

Feature-rich multi-agent framework with excellent plugin architecture, provider-based context injection, BM25 action filtering, and sandbox security — but sprawling complexity dilutes the core.

- **Core library:** 4 files, 507 lines total

Agents repeat mistakes. They fail the same way, miss optimization opportunities, and don't transfer successful strategies across tasks. This paper proposes a 4-component pipeline that automatically extracts structured, typed learnings from execution trajectories and injects them into future agent prompts via similarity-based retrieval.

ACT is a **two-stage RL training paradigm** that fixes the fundamental weakness of imitation learning (IL): IL teaches agents *what* expert actions look like, but never forces models to understand *why* those actions are better than alternatives.

OpenClaw-RL is a **live, online RL training framework** that trains language model agents *during production use* by extracting learning signals from the natural next-state feedback that already exists in every agentic interaction: user replies, tool outputs, error traces, test results, environment state changes.

open-multi-agent is a clean, well-typed TypeScript multi-agent framework that validates the coordinator-as-LLM pattern for auto-task-decomposition. Its architecture is lateral to Forge's SOP-driven execution — different paradigm, not better/worse.

> **Autoquant: a distributed quant research lab | v2.6.9**

ARIS is a prompt-engineering framework where the LLM *is* the runtime. 5 MCP servers bridge external LLMs for cross-model review, 5 CLI tools handle arxiv/Semantic Scholar fetching + GPU watchdog monitoring, and 49+ Markdown "skill" modules define composable research workflows (with YAML frontmatter) consumed directly by Claude Code/Codex/Cursor. The core architectural insight — "the LLM doesn't need a scheduler; the LLM IS the scheduler" — is orthogonal to Forge's programmatic orchestration but surfaces several quality patterns worth stealing: cross-provider adversarial review, provider-specific parameter clamping, thread history persistence for stateless APIs, and private dotfile API key fallback.

This is an empirical study, not a system. It analyzes pull requests from the **AIDev dataset** (HuggingFace), classifying reviewer compositions and evaluating signal quality of automated code review agent (CRA) outputs.

OpenAI_Agent_Swarm is an ambitious but under-implemented prototype. The vision documents describe a sophisticated hierarchical governance system (HAAS) that is almost entirely unbuilt. What IS built is a simple but instructive boss-worker queue system with one genuinely clever pattern: using OpenAI tool_call_ids as correlation tokens for cross-agent RPC.

The paper's **single core claim**: multi-agent system reliability is a **memory problem**, not a compute problem. The bottleneck in collaborative LLM agent systems looks "surprisingly familiar to computer architects" — it is the same memory hierarchy, bandwidth, and consistency challenge solved in CPUs, transplanted to semantic agent context.

This paper is a **diagnostic study**, not a system paper. It investigates a three-component agentic self-correction pipeline built on the Reflexion framework:

The paper addresses a real design question — **when to seek help** — but from the wrong direction for Forge. Our stack is LLM-native; MC Dropout cannot be applied to transformer inference.

### Rubric Scores

A **position/vision paper** (authors' own framing: *"we demonstrate the viability of integrating LM-Agents into FL with a proof-of-concept"*) proposing **Agentic-FL**: replacing static federated learning coordination protocols with LLM-based agents. The central argument is that existing FL solutions address isolated concerns (client selection, aggregation, privacy, communication) with fixed algorithms that cannot adapt to the dynamics of real FL environments. The paper claims LM-agents enable **holistic, simultaneous management** of all these concerns through contextual reasoning.