Python agents

让 AI 在安全约束内挖掘 A 股因子：一句话想法 → 受限 DSL 表达式 → 量化研究标准体检 → 策略回测（AFAC 金融智能创新大赛 · 量化投资方向参赛作品）

Enterprise Data Agent: NL → contract-bound DSL → deterministic SQL → DuckDB. StateGraph orchestration, sandboxed analysis, attribution skills, multi-provider gateway. LLM never emits raw SQL.

Autonomous local LLM agent with planning, tool execution, and self-reflection loop running fully offline

Self-contained bot that plays single-player skirmish matches in the browser game Chrono Divide (web Red Alert 2) using the game's official player API and a semantic decision model.

Provenance locked, three agent research harness for any AI coding agent. A Scout downloads sources, an Archivist hashes and quotes them, an Integrator cites them into your paper, report, or dataset. A stdlib Python engine verifies every quote, hash, and citation before the work can be marked done.

Skill orchestrator for Cursor, Claude Code & Codex. BM25 routing, editable subagent graphs, skill runbooks, and skills.sh discovery.

MCP-сервер: LLM работает на Авито через отдельный Chrome — поиск, объявления, кабинет и Pro, профили, правка форм; подтверждение необратимых действий

HEAMS: Holistic Evaluation of Agent Memory Systems. An open framework for evaluating memory systems for AI agents on performance and privacy.

基于 AgentScope 2.0 从零实现的企业级多智能体差旅助手

Literature retrieval from OpenAlex with Crossref source validation, producing screening tables for scoping reviews.

Equity valuation skills and data MCPs for DeepSeek Harness and Claude Code. SEC filings, bottom-up sector beta, sourced implied ERP.

飞书知识库 RAG Agent：增量同步 · 父子切块混合检索 · JWT 多用户隔离 · 微调图表专家（FastAPI + LangGraph + Milvus）

Tower defense game with a Gymnasium API for training AI agents (RL, heuristic, tournament).

A counter-espionage RL environment for scalable AI oversight, deception detection, sleeper-agent research, and reproducible security-first training.

[ASE 2026] ProvenanceGuard is a provenance-based runtime guardrail for LLM agent safety that detects misaligned tool calls before execution and keeps agent actions aligned with user intent.

P2Patch: exploit-guided vulnerability repair with LLM agents, and VulnRepairBench, a benchmark of 101 real CVEs with certified exploit suites.

Grocery agent for one household on one RTX 3090: NVIDIA Nemotron 3.5 Lightning via vLLM, traced in Arize Phoenix. Choose from what was retrieved, reject and count what was invented, every Yes/No is a label.

Autonomous multi-agent skill for AI agents. Cut token costs by 40–80% with structured workflows, quality gates, and isolated subagents.

Autonomous multi-agent software factory with parallel task dispatch, Kanban pipeline, isolated Git worktrees, and automated code review.

Agentic tool that distills a GBM into an interpretable GLM for Non-Life insurance pricing — an actuary reviews and approves every feature, grouping, interaction, and rating factor along the way.

Curated research on LLM-based social simulation, human behavior, and social interaction.

Code and evidence: process economics reorder catalyst rankings across computed and published catalysts (NH3 synthesis, CO2-to-methanol, Au/TiO2 control), with an agent that extracts literature data and runs the full chain

Two AI models build it; the harness checks the rules. fix/add/refactor/build/plan where tests are replayed against the original code — for Claude Code, Codex, Gemini CLI, Antigravity, OpenCode and Cursor. No API keys.

QueryNest is a research assistant that searches the live web and checks real-time weather, then answers with numbered sources

Lycosa — Open-source AI agent orchestration platform for building, coordinating, and deploying multi-agent LLM systems with RAG, MCP, knowledge bases, and automated workflows.

A production-grade OpenEnv-compliant reinforcement learning environment simulating a Security Operations Center (SOC) analyst workflow.

Replace an LLM agent's closed-set decisions with a cheap classifier trained from its own logs. Code for the paper TabAgent: Distilling LLM Agent Components into Cheap Classifiers.

Python prototype that chains GPT-4 prompts into Raven (interprets text-encoded patterns) and Seraph (ethical review) with an orchestrator and JSON-file memory. A prompt-orchestration experiment; it does not read or decode real neural signals.

LLM agents in place of rule agents across classic agent-based models; companion benchmark of SocioVerse2

Hachifish · 手机端 GUI Agent：说一句话，agent 通过官方无障碍通道在真机上把事办完。真实 App、不 root、零自有服务器（BYOK），完成前强制校验屏幕状态，随时可中止。

A concise repo for post-training LLM agents with SFT, on-policy distillation and RL fine-tuning on SWE-bench and multi-turn benchmarks.

GeoLingo: Translating Geometry into Language for Embodied Navigation. Metric depth and detections become text facts, and Jev picks each EB-Navigation action.

ChemCrow fork: local Ollama-backed reasoning LLM, custom-tool scaffold, Streamlit UI

Keep Claude Code's always-loaded MEMORY.md from silently truncating: budget guard, orphan coverage, verbatim domain folding into hubs. Deterministic, no LLM for the mechanical parts.

This repository contains the full demos, code, and materials used in my DevFest Istanbul Türkiye 2025 session. The talk covers two major parts: Part I — Building Agentic AI Workflows with xAI on Vertex AI Part II — Rebuilding the Same Agent Using Vertex AI Agent Builder

SKILL.state long-horizon agent-skill runtime for Hermes Agent (arXiv 2608.26263) — explicit-state execution, O(1) prompt, OpenAI-compatible backends

No Promises Without Proof. A dependency-free discipline that forces agents to pair every success claim with evidence and a proof state (OBSERVED/INFERRED/UNVERIFIED/BLOCKED), then derives the completion verdict deterministically.

Analytics-engineering warehouse (dbt, DuckDB, MotherDuck) around a frozen NBA player-stat model: contracts, reconciliation to committed artifacts, a 23-feature SQL parity proof, lineage ending in exposures, and a pre-registered out-of-sample paper evaluation against market lines. Free tiers only.

A verified browser agent: a cheap Jev guard (evidence-checked completions, a destructive gate) in front of any LLM browser driver, with Jev taking the mechanical steps in dual mode. Built on browser-use; every number pre-registered and measured.

You can work for Wall Street. Agent-agnostic equity-research pipeline: methodology distilled from 100 real sell-side reports, 10 stages / 8 roles, real dual-market data (US + A-share, no API key), 5 institutional chart families, 6 anonymised sell-side layout templates, Markdown → Word/PDF.

Hooks and a ledger CLI that make an AI-assisted smart-contract audit follow its own rules: hypotheses are killed by failing PoCs, not prose.

Enterprise AI Operations Platform for orchestrating AI agents, workflows, metrics, and execution monitoring.

Switchyard: a small local daemon that keeps one live session per coding agent (agy, Claude Code, OpenCode, Codex). Spawn, watch, steer, interrupt and stop them from one CLI, MCP bridge or dashboard, with tier routing and a policy guard.

The Experience Lifecycle of lifelong, self-evolving & memory-augmented LLM agents — acquisition, representation, retrieval, consolidation, abstraction, internalization, revision, distribution + governance. Companion to the survey.

Integrity scenarios for agent runtimes, distilled from real fleet incidents — deterministic fault-injection checks against real SDKs

Contact angle parameter search pipeline: COMSOL+MATLAB co-simulation with Bayesian/GA/PSO/CMA-ES optimizers

YOLO real-time object detection dashboard with MCP integration

Autonomous Tier-1 Equity Research & DCF Financial Modeling Agent. Produces 10-tab institutional Excel models, 16-page PDF initiation reports, and multi-model valuation.

Pediatric dose-extrapolation agent: a defensible starting dose for a child from adult pharmacokinetics via allometry × organ maturation, with a cited, graded, auditable rationale. Multi-agent (Opus + Sonnet, live PubMed/openFDA). Decision support, not prescribing.

Experiments in a small RL environment: do LLM agents game the rules when the answer is within reach, and can you tell them not to? Claude Opus 5, Qwen 3.8 27B, abliterated Qwen.

When Does Query Expansion Help Agent Memory? A Multi-Seed Study of Selective Expansion and Gate Reliability. Code, results and paper.

Empirical case study of measurement validity in LLM-agent prompt-injection defense evaluation: how success endpoints, labeling rules and reporting units change conclusions on the same traces.

Open RDelta/Babel Index framework for LLM agents and world-systems: detect context degeneration, agent loops, tool fragility, and collapse with NEOTH.

LLM swarm simulation of segmented market reactions — 120 agents with personas, memory, and herding, validated against the 2023 SVB collapse

Measure your Claude Code skills and rules, and keep only the ones that help. The tool, its evaluation lab and data for the paper arXiv:2610.14216.

Reproducible simulations of LLM agents, evidence provenance, and institutional mechanisms for cooperation under mixed incentives.

A Tailnet-only Misskey lab where ten persona-driven Hermes agents build ongoing conversations.

Citation-grounded RAG agent over arXiv papers: section/table-aware chunking, vector search + cross-encoder re-ranking, FAISS, tool-calling agent, groundedness check and an evaluation harness.

NeurIPS 2026 | When does stored evidence stop being usable? Official code and evaluation artifacts for testing agent-memory reliability as history grows under explicit memory-call budgets.

Agentic framework for hardware-aware, layer-wise LLM quantization using sequential optimization and on-device benchmarking.

MCP server for SEC EDGAR filings, built for LLM agents: typed outputs, section extraction, XBRL normalization, filing-to-filing diffs.

Official implementation of "RefP2C: Reflective Paper-to-Code Development Enabled by Fine-Grained Verification".

A skill that lets an LLM agent untangle the latent space of a frozen neural audio codec by experiment, then use that map to produce sounds the user asks for.

An ML platform for work-queue risk, built on the GitHub public event firehose. Point-in-time-correct feature store, PySpark medallion, MLflow serving, and a bounded LLM agent whose every claim is mechanically verified against its own tool calls.

[arXiv] ExpHarness: Model-Agnostic Experience Learning through a Trainable Harness

IDE skills for scaffolding AI agent harnesses: ReAct, plan-and-execute, self-evolving, reflexion, multi-agent, code agent, RAG memory

We propose a simple, task-, LLM-, and harness-agnostic method to curate high-performing initial discovery populations which can be used to successfully warm-start and boost the performance of LLM-driven discovery.

An evidence-driven map of Recursive Self-Improvement: systems, scorecards, benchmarks, research papers and live project radar.

Typed-judgment model (TypeSafe Jev) driving task-level decisions on a simulated xArm7: reliability measurements, failure map, sensor-freshness gate

基于 tau2-bench Retail 的电商多步 Agent 后训练与可靠性评测：教师数据、多步 SFT、Agentic GRPO、规则验证与轨迹诊断。

Run many Claude Code sessions in parallel against one shared task queue — concurrency derived from the work and from what your machine can actually spare. 100 tasks never means 100 browser tabs.

Evaluation framework for A/B test analysis agent.

Your scheduled job says exit 0 — prove it did the work. Three stdlib-only checks: output freshness, silent no-op detection, rollout proof by reading the fact back. Sanitized from a live agent fleet. Free, MIT.

Composable, stateful AI workflows built with LangGraph, implementing sequential, parallel, conditional, iterative, human-in-the-loop, and multi-agent orchestration through graph-based execution. Models use nodes, edges, schema-defined state, routing, checkpointing, tool orchestration, LLM-driven control flow, and graph compilation.

Enterprise-grade AI Data Loss Prevention (DLP) for LLM agents using Just-In-Time encryption and format-preserving tokenization.

Curated awesome list of open source agent frameworks and tools for building, running, connecting, and observing AI agents.

图解 Agent 系统：机制、开源实现与可编辑架构图

Multi-agent incident response with LangGraph: parallel investigators, cited evidence, prompt-injection quarantine, and a human approval gate.

QED Proof Python SDK (pip install qed-proof): submit an AI agent's claimed work, get an independently verified verdict, and verify the signed receipt offline.

An asynchronous social sandbox where humans and LLM agents share persistent characters in an emergent ECS simulation.

A hierarchical multi-agent deep research system built with LangGraph, Tavily, evaluator, and red-team workflows.

Market-Language-Action (MLA) agent for explainable stock trading using LoRA-tuned Qwen models, teacher–student distillation, and out-of-sample backtesting

Local-first Markdown vault blueprint for multi-agent memory, session claims, and safe handoffs (Codex/Claude-friendly).

gguf apex custom-quantization unsloth-studio moe coding

gradio region:us

gradio mcp-server region:us

gradio region:us

gradio trackio region:us

gradio region:us

gradio mcp-server region:us

static trackio trackio-logbook open-experiment icml2026-repro paper-tPcL3fgkjc region:us

Crewai workflow example (illustrated e-book)

A human-in-the-loop policy guardrail for LLM / AI-agent actions — approve, escalate, or reject with an auditable reason and risk score. Zero deps, typed, framework-agnostic (CrewAI/LangGraph).

AI agent crew that generates hyper-local neighborhood guides with real estate data, crime stats, and lifestyle insights using CrewAI + 13 MCP servers + Google Gemini.

Python SDK and CLI for Vulnify: check an AI agent action and get ALLOW, REVIEW or BLOCK before it runs. Adapters for CrewAI and LangGraph.

AI-powered content creation framework that automates trend research, writing, and SEO. A team of specialized agents (analyst, writer, optimizer) generates high-quality, platform-specific content (blogs, LinkedIn, Instagram, etc.) based on real-time trends.

AgentForge is a hands-on platform for building, testing, and comparing modern AI agent frameworks. It features real-world examples, multi-agent workflows, and benchmarking tools to evaluate performance, scalability, and design trade-offs across leading solutions.

Open-source AI EDR that detects and stops ransomware, Cobalt Strike, and living-off-the-land attacks in real time — with a beautiful web dashboard and ChatGPT-style AI analyst you can talk to.

Autonomous workforce planning system using XGBoost, LangGraph, guardrails, and supervisor-led VET/VTO staffing recommendations with traceable decision support.