Skip to content

agentic frameworks

Jump to bottom
Nicolas Cravino edited this page Jul 22, 2026 · 9 revisions

id: agentic-frameworks title: Agentic Frameworks tags: [python, langgraph, langchain, azure, openai, mlx, audio, cli, stt, tts, rag, benchmark, vllm, cuda] created: 2026-04-11 updated: 2026-07-22 freshness: fresh

Agentic Frameworks

Summary

Nine repositories demonstrating diverse agentic AI patterns — from a mental model visualization (Agent Stack) through workflow conversion (N8n2LangGraph), coding assistants (DeepAgent Azure CLI on Azure/MLX, RalphDGX on DGX Spark/vLLM), voice analysis research loops (SST-AutoResearch), experimental multi-agent requirements generation (AutoRequirementsAgent), and a head-to-head benchmark of agentic-RAG vs three retrieval baselines (Wiki-vs-RAG). LangGraph and AutoGen are the dominant orchestration frameworks; tool-loop budgets and Karpathy-style autoresearch loops appear repeatedly.

Key claims

  • Agent Stack decomposes agentic systems into 10 layers, reframing the LLM as a fallible component within a larger deterministic system. ^src:agent-stack

  • DeepAgent Azure CLI provides a full coding agent (file R/W, shell, glob, grep, sub-agents) wired to Azure OpenAI with Textual TUI, reasoning effort control (low/medium/high), and SQLite session persistence. ^src:deepagent-azure-cli

  • DeepAgent Azure CLI can run fully offline using a local MLX server that emulates the Azure OpenAI API. ^src:deepagent-azure-cli

  • N8n2LangGraph converts n8n workflow JSON into standalone LangGraph Python scripts, with optional reflection loops and LLM-assisted translation for unsupported nodes. ^src:n8n2langraph

  • SST-AutoResearch implements a Karpathy-style autoresearch loop where the LLM drives hypothesis → design → execute → evaluate → reflect cycles over speaker voice dynamics (Takens' embedding, Lyapunov exponents, recurrence analysis). ^src:sst-autoresearch

  • RalphDGX wraps the DeepAgents Code CLI (dcode) in a Ralph/tenacity goal loop on DGX Spark: an install → patch → overlay harness that retargets DeepAgents at a local vLLM server (Qwen3-Coder-Next-FP8, :8000/v1), with a doctor/install/test/run lifecycle and cloud fallback. ^src:ralph-dgx

  • AutoRequirementsAgent experiments with strict JSON-format inter-agent messaging and state-transition functions, using nested GroupChat with local Ollama LLMs (Nemotron 70B, Hermes 70B) for deterministic multi-agent coordination. ^src:autogen-requirements-agent

  • Wiki-vs-RAG benchmarks four agent patterns over the sw30labs wiki corpus (single-shot RAG, agentic-RAG with ≤8-turn tool loop, wiki navigation with precomputed sitemap, and QMD local hybrid retriever) with Claude Sonnet 4.6 as the agent and Claude Opus 4.7 as the LLM-judge. ^src:wiki-vs-rag

  • Agentic-RAG is the Pareto winner on the 30-query eval — same quality as wiki-nav (4.77 vs 4.69, Δ inside noise) at ~45% of the per-query cost; beats single-shot RAG 11–2 with 17 ties for 4.2× the spend. ^src:wiki-vs-rag

  • QMD (BM25 FTS5 + EmbeddingGemma + Qwen3 reranker, fully on-device) is the cheapest arm at $0.005/query with zero embedding API spend, ties on narrow factoids, and collapses on branch / cross-branch synthesis (quality 2.83 / 3.37 vs 4.5+ for the other three) — confirming its "AI grep" framing. ^src:wiki-vs-rag

  • LangGraph Checkpoints vs Stores is a runnable, offline reference that distinguishes thread-scoped checkpoints (resume, time travel, interrupts, fault tolerance) from cross-thread stores (durable user memory), with every quoted output captured from a real CI-gated run and no LLM calls. ^src:langgraph-checkpoints-vs-stores

  • Chapters 2–3 of the checkpoints-vs-stores reference take the same graphs to production persistence — kill-and-resume across processes, a one-graph backend matrix over memory/SQLite/Postgres/Redis with CI against real service containers — and to human-in-the-loop interrupts and time travel. ^src:langgraph-checkpoints-vs-stores

  • Venture Pathfinder scans local git repos and research assets into per-project profiles in a Neo4j graph, then uses a LangGraph engine and a Fabric-inspired Pattern Engine (driven by a local OMLX model) to surface recurring patterns and white-space venture paths across a body of work. ^src:venture-pathfinder

Open questions

  • Could N8n2LangGraph be used to convert workflows into DeepAgent Azure CLI custom tools?
  • Could Wiki-vs-RAG's tool-loop harness be repurposed to benchmark DeepAgent Azure CLI's agent loop against the same corpus?
  • Does the Wiki-vs-RAG result generalize to corpora that aren't already organized as a Karpathy-style wiki?

Related

Articles

Sources

Home | Index | Sitemap | Sitemap-Stacks | SCHEMA | Log

Clone this wiki locally