Desk
Papers, model cards, and lab releases from the last two weeks. The library below is a fixed reading list, not the live feed.
Papers and models come from the lab feeds. Code releases are whatever shipped on GitHub this week, including repos nobody added by name.
Microsoft
Policy enforcement, zero-trust identity, execution sandboxing, and reliability engineering for autonomous agents. Intercepts tool calls, messages, and delegations in deterministic code before they hit the wire. Covers 10/10 OWASP Agentic Top 10.
Agent AI
58
Company papers
103
Companies
53
Showing
58
Curated canon — 58 items. Use search and company filters above for live source intel.
Yuxing Lu, Yicheng Chen, Shanchan Wu, Sercan Ö. Arik
A Procedural Graph organizes agent how-to knowledge as editable (procedure, relation, procedure) triplets. At each step it localizes the active node and turns the local subgraph into situational guidance, then self-evolves from failed vs successful trajectories.
Shanghai AI Lab
Shanghai AI Lab
Shanghai AI Lab (May 2026). Lightweight and scalable alignment framework for agent safety.
ServiceNow
ServiceNow
Central hub to discover, secure, govern, observe, and measure any AI across the enterprise. 30+ enterprise integrations. Gartner MQ Leader.
IBM
IBM
Enterprise AI assurance and GRC platform. Governance Graph mapping AI assets, policies, risks. 200+ regulatory frameworks. Gartner MQ Leader.
LangChain
LangChain
Category reference point. Framework-agnostic tracing, OpenTelemetry, LLM-as-judge, prompt management, agent deployment. Official LangSmith observability docs.
Arize
Arize
Open-source observability. OpenTelemetry-based tracing, embedding clustering, HDBSCAN, UMAP visualizations. Self-hostable.
Langfuse
Langfuse
MIT-licensed tracing, prompt management, LLM-as-judge, CI/CD quality gates. Acquired by ClickHouse (Jan 2026). ~31.2k GitHub stars.
Comet
Comet
Apache 2.0 LLM observability and evaluation. End-to-end tracing, experiment tracking, PyTest integration, automated agent optimizer.
NVIDIA
NVIDIA
Programmable guardrails for LLM apps. Topic control, PII detection, RAG grounding, jailbreak detection. Apache 2.0. ~6,768 GitHub stars.
Meta
Meta
System-level guardrail framework for AI agents. Prompt Guard 2, Agent Alignment Checks, code security. Part of PurpleLlama.
Guardrails AI
Guardrails AI
Schema-driven LLM output validation and correction. Developer-first. Apache 2.0. ~7,190 GitHub stars.
Protect AI
Protect AI
35+ input/output scanners. Prompt injection, PII, secrets, topic banning. MIT licensed.
Amazon
Amazon
Managed content filters, denied-topics, PII redaction inline with Bedrock inference.
Microsoft
Microsoft
Harm scoring (hate, violence, self-harm, sexual). Prompt Shields for jailbreak detection.
Configurable safety-attribute scoring. Checks Guardrails API. GuardGemma safety models.
Traceloop
Traceloop
OpenTelemetry-based observability for LLM applications. Official OSS repo; now part of ServiceNow AI Control Tower.
CrowdStrike
CrowdStrike
AI Detection and Response. Runtime visibility and control for AI agents. AI gateway. Supports NeMo Guardrails.
Wiz
Wiz
AI Security Posture Management. Code-to-runtime security graph. Autonomous security workflows.
NVIDIA
NVIDIA
LLM vulnerability scanner. Probes for hallucination, data leakage, prompt injection, bias.
Promptfoo
Promptfoo
Open-source red teaming for 50+ vulnerability types. OWASP, NIST, EU AI Act mapping. Official repo.
Shanghai AI Lab
Shanghai AI Lab
Shanghai AI Lab (Jan 2026). Diagnostic guardrail framework with three-dimensional taxonomy for risky agent behaviors.
ByteDance
Pengfei Gao, Zhao Tian, Xiangxin Meng, and colleagues
Ensemble coding agent with generate/prune/select modules for repo-level issues. First on SWE-bench Verified at 75.20% Pass@1; code at github.com/bytedance/trae-agent.
OpenAI
OpenAI
Open-source coding agent that runs locally in the terminal, with sandboxed execution and approval modes. Official repository.
Open-source, code-first toolkit for building, evaluating, and deploying multi-agent systems. Model- and deployment-agnostic, optimized for Gemini, with tools, MCP, workflows, and Vertex AI Agent Engine.
A2A Project (Linux Foundation)
Open agent-to-agent interoperability protocol: agent cards, tasks, and streaming messages so agents from different vendors can delegate and collaborate.
Cohere
Team Cohere
Agent-optimised enterprise model with 23 business languages, hybrid architecture, grounded RAG, and tool use. Open research weights for Command A and Command R7B.
Meta
Meta AI
Meta AI (Apr 2025). Open-source guardrail system for secure AI agents. Prompt Guard 2, Agent Alignment Checks.
OpenAI
OpenAI
Lightweight production framework for multi-agent workflows: agents, tools, handoffs, guardrails, sessions, and tracing. Provider-agnostic successor to Swarm.
Anthropic
Anthropic
Anthropic's agentic coding tool for the terminal: issues, docs, and reference for the CLI, hooks, MCP integration, and SDK. Official repository.
Hugging Face
Hugging Face
Minimal agent library where agents write their actions as Python code, with sandboxed execution and Hub-hosted tools. Official repository.
Oxford
Sumeet Ramesh Motwani, Chandler Smith, Rocktim Jyoti Das, and colleagues
Oxford multi-agent post-training: generator, verifier, and refiner specialize via value iteration on search trees. +15.66% relative on MATH vs the same base LLM.
Anthropic
Model Context Protocol maintainers
Open protocol for connecting LLM applications to tools, resources, and prompts over JSON-RPC. Spec source plus schema; the wire format most agent frameworks now speak.
Microsoft
Adam Fourney, Gagan Bansal, Hussein Mozannar, and colleagues
AutoGen-based generalist team with an Orchestrator plus WebSurfer, FileSurfer, Coder, and ComputerTerminal agents. Modular multi-agent design for web, files, and code tasks; code released as Magentic-One.
Cisco
Cisco
Built on Robust Intelligence acquisition (Sep 2024). AI security category pioneer. Cisco Foundation AI core.
Sierra
Shunyu Yao, Noah Shinn, Pedram Razavi, Karthik Narasimhan
Simulated users plus domain policies for retail and airline support. Measures whether agents reach the correct end state and how consistently (pass^k) — even GPT-4o passes under 50%.
Cohere
Cohere
Deployable enterprise agents with retrieval, rerank, and tool use: chat, search, and RAG starters wired to Command models.
Princeton
John Yang, Carlos E. Jimenez, Alexander Wettig, and colleagues
Custom agent-computer interfaces for editing, navigation, and tests. SOTA on SWE-bench and HumanEvalFix — the interface-design paper behind practical coding agents.
HKU
Tianbao Xie, Danyang Zhang, Jixuan Chen, and colleagues
369 real computer tasks on Ubuntu, Windows, and macOS with execution-based evaluation. Humans 72%; best model 12% — the standard computer-use agent benchmark.
ETH Zurich
Shenglai Zeng, Jiankun Zhang, Pengfei He, and colleagues
First systematic RAG privacy audit: novel attacks leak the retrieval database, yet RAG also mitigates training-data leakage. Essential for agents on proprietary data.
LangChain
LangChain
Low-level orchestration framework for stateful, long-running agents: graphs, persistence, human-in-the-loop, and streaming. Official repository.
Hugging Face
Gregoire Mialon, Clementine Fourrier, Craig Swift, and colleagues
466 real-world questions needing reasoning, multimodality, browsing, and tool use. Humans 92%, GPT-4 with plugins 15% — the benchmark that humbled every generalist agent.
University of Washington
Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, Hannaneh Hajishirzi
The model decides when to retrieve and emits reflection tokens that critique its own passages and output. Outperforms ChatGPT and retrieval-augmented Llama2-chat on open-domain QA and fact verification.
Princeton
Carlos E. Jimenez, John Yang, Alexander Wettig, and colleagues
2,294 real GitHub issues across 12 Python repos — the benchmark every coding agent from Claude Code to Codex is measured against.
Microsoft
Qingyun Wu, Gagan Bansal, Jieyu Zhang, and colleagues
Microsoft Research framework for multi-agent conversation: customizable, conversable agents that cooperate via dialogue, with human participation and tool use. Foundation for later Magentic-One work.
Tsinghua
Xiao Liu, Hao Yu, Hanchen Zhang, and colleagues
Eight interactive environments — OS, database, knowledge graph, card game, puzzles, household, web shopping, web browsing — for multi-turn agent evaluation across 27+ LLMs.
CMU
Shuyan Zhou, Frank F. Xu, Hao Zhu, and colleagues
Self-hosted, fully functional websites (e-commerce, forums, GitLab, CMS) with 812 long-horizon tasks. Best GPT-4 agent reached 14.41% vs 78.24% for humans.
NVIDIA
Guanzhi Wang, Yuqi Xie, Yunfan Jiang, and colleagues
First LLM-powered lifelong learning agent in Minecraft: automatic curriculum, an ever-growing skill library of executable code, and iterative prompting from environment feedback.
Berkeley
Shishir G. Patil, Tianjun Zhang, Xin Wang, Joseph E. Gonzalez
Fine-tuned LLaMA that writes correct API calls across 1,600+ APIs, with retriever-aware training that cuts hallucinated calls. Origin of the Berkeley Function-Calling Leaderboard.
Princeton
Shunyu Yao, Dian Yu, Jeffrey Zhao, and colleagues
Search over thought units with self-evaluation and backtracking. GPT-4 jumps from 4% to 74% on Game of 24.
Stanford
Joon Sung Park, Joseph O'Brien, Carrie Cai, and colleagues
Twenty-five LLM agents with memory, reflection, and planning populate a Sims-style town — emergent parties, dates, and coordination from one user notion.
Microsoft
Yongliang Shen, Kaitao Song, Xu Tan, and colleagues
An LLM controller plans tasks, selects expert models from the Hub, executes them, and summarizes results. Early blueprint for model-orchestrating agents (code: microsoft/JARVIS).
MIT
Noah Shinn, Federico Cassano, Edward Berman, and colleagues
Reinforces agents through linguistic feedback, not weight updates: verbal self-reflection in episodic memory. 91% pass@1 on HumanEval.
Microsoft
Microsoft
Model-agnostic SDK for agents and plugins in C#, Python, and Java, with an Agent Framework and process orchestration. Official repository.
Meta
Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, and colleagues
Meta FAIR shows a language model can teach itself when to call APIs — calculator, Q&A, search, translation, calendar — via self-supervised tool-use annotations, with little extra training cost.
Stanford
Stanford NLP
Programming — not prompting — language models: declarative modules with optimizers that compile pipelines into effective prompts and weights.
Shunyu Yao, Jeffrey Zhao, Dian Yu, and colleagues
Interleaving reasoning traces with actions lets LLMs plan and pull in external knowledge. Beats SOTA on HotpotQA, FEVER, ALFWorld, and WebShop.
Meta
Patrick Lewis, Ethan Perez, Aleksandra Piktus, and colleagues
The original RAG recipe: parametric seq2seq plus a dense Wikipedia index. Set SOTA on three open-domain QA tasks and launched the retrieval ecosystem agents still rely on.