#agents

20 Oct 2026 · Study
Agentic AI · Agent safety
The main risks when agents act (prompt injection through tool outputs, excessive permissions, unconfirmed side effects) and simple defences: separating data from instructions, allow-lists, argument validation and human confirmation.
9 Oct 2026 · Study
Agentic AI · Agent evaluation
Evaluating agents means checking the final answer and the trajectory: task success, tool-call correctness, wasted steps and cost, scored here over a few recorded agent runs.
7 Oct 2026 · Projects
Built this: AgentForge
A three-agent CrewAI research assistant (Researcher, Analyst, Reviewer) whose sources come only from real tool activity and whose failures are reported as failures.
5 Oct 2026 · Paper summary
Paper summary: RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations
RealCompanion releases ten real long-term human–AI companion relationships (27,218 messages over up to 120 days) and finds that people rarely refer back, that AI can't tell when the past matters, and that it reads more into people than they revealed.
30 Sept 2026 · Study
Agentic AI · Model Context Protocol (MCP)
MCP standardises how applications expose tools, resources and prompts to LLM clients; an MCP server built with the official Python SDK and queried from an in-process client.
30 Sept 2026 · Paper summary
Paper summary: Raven: The Harness of Harnesses for Composable Agentic Intelligence
Raven is an open-source multi-agent ecosystem that automatically builds and evolves specialised model–harness pairs and orchestrates them across domains through a Host Agent, with stored experience turned into reusable skills.
28 Sept 2026 · AI updates
Hugging Face Hub adds RL environments
The Hugging Face Hub gained a filter for RL environments: an environment is a dataset repo (tasks, tests, containers and a reward rule) with a "Use this dataset" command for frameworks like Harbor, Verifiers and NVIDIA NeMo Gym.
22 Sept 2026 · Paper summary
Paper summary: RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
Automatically evolving an agent's harness (prompts, control flow, tools, memory) tends to overfit the tasks it evolves on; RRSI regularises the proposal and selection steps and keeps gains on out-of-distribution benchmarks.
21 Sept 2026 · Study
Agentic AI · Multi-agent systems
Multi-agent patterns (supervisor/router, sequential crew, validator) with a small supervisor that routes questions to specialist agents, and how my Multi-Agent-RAG and AgentForge projects use them.
15 Sept 2026 · Paper summary
Paper summary: Atria Dawn: The Dawn of Agentic Superintelligence
Atria Dawn Preview is an agentic language model for research and engineering trained on verified tool-mediated experience; alongside benchmarks, the paper studies 769 real task records from 56 people working with it.
10 Sept 2026 · Study
Agentic AI · Agent memory
Short-term memory (the conversation in the context window) versus long-term memory (facts stored outside and retrieved when relevant), with a small long-term store built on embeddings.
10 Sept 2026 · Paper summary
Paper summary: Scaling Automatic Research Agents via World Models
RL for automatic-research agents is bottlenecked by slow, sandboxed experiment execution; World Model RL replaces execution with a learned world model plus debiasing and denoising, training 3–4× faster.
1 Sept 2026 · Study
Agentic AI · Planning
Plan-and-execute: write a plan of steps with dependencies first, then run them in order, re-planning when a step fails, implemented as a small task graph executor.
21 Aug 2026 · Study
Agentic AI · Tool and function calling
Structured tool calling with a real local model: Python functions become JSON schemas, Qwen2.5-0.5B-Instruct emits tool calls through its chat template, and the app parses and executes them, including one wrong call.
13 Aug 2026 · AI updates
Gemini 3.7 Flash: three weeks after 3.6
Google released Gemini 3.7 Flash three weeks after 3.6 Flash, reporting large gains on coding (DeepSWE v1.1 65.3% vs 49.0%), web development and document-heavy knowledge work, at half 3.6 Flash's introductory price.
12 Aug 2026 · Study
Agentic AI · ReAct
ReAct interleaves reasoning ("Thought") with tool use ("Action") and results ("Observation"); the prompt format, a parser for model output, and a full trace with a scripted model.
3 Aug 2026 · Study
Agentic AI · What is an agent
An LLM agent is a loop: the model decides an action, a tool runs it, the result goes back to the model, until it decides it's done; the loop written in plain Python with a scripted policy.
28 Jul 2026 · Paper summary
Paper summary: Kimi K3: Open Frontier Intelligence
Kimi K3 is an open-weight 2.8T-parameter MoE (104B active) with native vision and a 1M-token context, reported as about 2.5× more scaling-efficient than Kimi K2, though still behind the strongest proprietary models.
15 Jul 2026 · Paper summary
Paper summary: Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models
The authors note that an agent's act–observe–continue loop has the same shape as a function call site, and use function-level fill-in-the-middle mid-training on Python repos to improve coding agents on SWE-Bench.
24 Jun 2026 · Paper summary
Paper summary: Qwen-AgentWorld: Language World Models for General Agents
Qwen-AgentWorld models are language world models that predict how agent environments respond across 7 domains, trained on 10M+ real interaction trajectories, and are reported to outperform frontier models at simulation on a new benchmark.
16 Jun 2026 · Paper summary
Paper summary: JoyAI-VL-Interaction: Real-Time Vision-Language Interaction Intelligence
JoyAI-VL-Interaction is an open 8B vision-language model that watches a live video stream and decides each second whether to stay silent, respond, or hand off to a bigger background model.
19 May 2026 · AI updates
Gemini 3.5 Flash: agents and coding at Flash speed
Google DeepMind introduced Gemini 3.5 with 3.5 Flash first, reporting it beats Gemini 3.1 Pro on agentic and coding benchmarks (Terminal-Bench 2.1 76.2%, MCP Atlas 83.6%) at about 4× the output speed of other frontier models.
19 May 2026 · Study
LangChain · LangGraph
LangGraph models an application as a state machine: typed state, nodes that update it, and conditional edges, used here to build a small corrective-retrieval loop that rewrites the query when retrieval looks weak.
19 May 2026 · Paper summary
Paper summary: Code as Agent Harness
A survey that treats code as the infrastructure of agents, not just their output: the interface agents act through, the mechanism for planning, memory and verification, and the shared artefact that coordinates multiple agents.
29 Apr 2026 · Paper summary
Paper summary: Recursive Multi-Agent Systems
RecursiveMAS treats a whole multi-agent system as one recursive computation in latent space, linking agents with a light RecursiveLink module and training them jointly; it reports +8.3% accuracy and 1.2–2.4× faster inference.
24 Apr 2026 · Projects
Built this: Orbit-Wars strategy agent
A rule-based strategy agent for the Kaggle Orbit Wars environment using predictive intercepts, economic target scoring, defensive "fortress" logic and sun-avoidance manoeuvres.
20 Apr 2026 · Projects
Built this: Enterprise Agentic RAG Orchestrator (Multi-Agent-RAG)
A Supervisor → Worker → Validator multi-agent RAG system with LangGraph Corrective RAG, hybrid dense + BM25 retrieval with cross-encoder reranking, a Qdrant semantic cache and a security layer.
12 Mar 2026 · Paper summary
Paper summary: OpenClaw-RL: Train Any Agent Simply by Talking
OpenClaw-RL turns the next state after every agent action (a user's reply, a tool output, a GUI change) into live training signal, so a personal agent improves online simply by being used.
18 Feb 2026 · Paper summary
Paper summary: GLM-5: from Vibe Coding to Agentic Engineering
GLM-5 targets "agentic engineering" beyond vibe coding, using DSA attention to cut long-context cost and an asynchronous RL infrastructure plus new agent RL algorithms for long-horizon tasks.
22 Jan 2026 · Paper summary
Paper summary: A Survey of Agentic Reasoning for Large Language Models: Towards Recursively Self-Improving and Collective Agents
A survey organising agentic reasoning into three layers (foundational single-agent skills, self-evolving agents, collective multi-agent reasoning), each split into in-context and post-training approaches.
1 Jan 2026 · Paper summary
Paper summary: Youtu-LLM: Unlocking the Native Agentic Potential for Lightweight Large Language Models
Youtu-LLM is a 1.96B model pre-trained from scratch (not distilled) with a commonsense → STEM → agent curriculum over about 11T tokens and agentic mid-training, reported as state of the art below 2B parameters.
2 Dec 2025 · Paper summary
Paper summary: From Code Foundation Models to Agents and Applications: A Comprehensive Survey and Practical Guide to Code Intelligence
A large survey and practical guide to code LLMs, from data curation and pre-training through SFT, RL and coding agents, comparing general and code-specialised models and the gap between benchmarks and real development.
25 Nov 2025 · Paper summary
Paper summary: General Agentic Memory Via Deep Research
General Agentic Memory keeps a light memory plus the full history in a page-store, then uses a "researcher" agent to build the right context at request time, like just-in-time compilation for memory.
13 Nov 2025 · Paper summary
Paper summary: Lumine: An Open Recipe for Building Generalist Agents in 3D Open Worlds
Lumine is an open recipe for a generalist game-playing agent built on a vision-language model: pixels in at 5 Hz, keyboard and mouse out at 30 Hz, completing a five-hour storyline and transferring zero-shot to other games.
17 Oct 2025 · AI updates
LangChain 1.0 released
LangChain reached 1.0. The release notes list the final changes before it: new middleware (shell, file search, tool retry), an injectable ToolRuntime, state-schema options for create_agent, and Python 3.14 support.
11 Aug 2025 · Paper summary
Paper summary: GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models
GLM-4.5 is an open MoE LLM (355B total, 32B active) with hybrid thinking and direct modes, trained on 23T tokens plus expert iteration and RL, reporting 70.1% TAU-Bench, 91.0% AIME 24 and 64.2% SWE-bench Verified.
29 Jul 2025 · Paper summary
Paper summary: Agentic Reinforced Policy Optimization
ARPO is an RL algorithm for multi-turn tool-using agents that samples extra rollouts right after tool calls, where the model's token entropy spikes, and is reported to beat trajectory-level RL with half the tool calls.
18 Jul 2025 · Paper summary
Paper summary: A Survey of Context Engineering for Large Language Models
A survey of over 1,400 papers that defines "context engineering": retrieving, processing and managing everything put into an LLM's context, and how RAG, memory, tools and multi-agent systems build on it.
8 Jul 2025 · Paper summary
Paper summary: MemOS: A Memory OS for AI System
MemOS proposes a memory operating system for LLMs that manages plaintext, activation-based and parameter-level memory as one schedulable resource, built around "MemCube" units with provenance and versioning.
13 May 2025 · Paper summary
Paper summary: Seed1.5-VL Technical Report
ByteDance's Seed1.5-VL pairs a 532M vision encoder with a 20B-active MoE LLM and reports state of the art on 38 of 60 public benchmarks, plus strong GUI-control and gameplay agent results.
4 Apr 2025 · Paper summary
Paper summary: Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems
A book-length survey that frames LLM agents through brain-inspired modules (memory, world model, reward, goals, emotion), then covers self-improvement, multi-agent systems and agent safety.
7 Mar 2025 · Paper summary
Paper summary: START: Self-taught Reasoner with Tools
START teaches a long chain-of-thought reasoning model to call a Python interpreter mid-reasoning, using inserted hints ("Wait, maybe using Python here is a good idea") and rejection-sampling fine-tuning.
20 Feb 2025 · Paper summary
Paper summary: Qwen2.5-VL Technical Report
Qwen2.5-VL adds precise object localisation, robust document parsing and hour-long video understanding through a natively dynamic-resolution ViT and absolute time encoding; the 72B model is reported to match GPT-4o and Claude 3.5 Sonnet.
4 Feb 2025 · AI updates
Open-source DeepResearch in 24 hours
The day after OpenAI's Deep Research launch, Hugging Face set out to reproduce its agentic framework openly in 24 hours, built on smolagents' CodeAgent and evaluated on GAIA.
31 Dec 2024 · AI updates
smolagents: agents that write their actions as code
Hugging Face launched smolagents, a minimal agent library whose main agent type writes its actions as Python code, plus a clear way to think about "agency" as a spectrum.
27 Nov 2024 · Paper summary
Paper summary: ShowUI: One Vision-Language-Action Model for GUI Visual Agent
ShowUI is a vision-language-action model for GUI agents that sees screenshots like a person, trims redundant screen tokens with a UI graph, and interleaves vision, language and actions.
13 Aug 2024 · Paper summary
Paper summary: The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
The AI Scientist chains LLM steps that generate research ideas, write and run experiment code, write a full paper and review it with an automated reviewer, as a framework for fully automated research.
7 Aug 2024 · AI updates
LangGraph 0.2: pluggable checkpointers for durable agents
LangGraph v0.2 split state persistence into dedicated checkpointer libraries (in-memory, SQLite and an open-sourced Postgres version) for session memory, error recovery and human-in-the-loop.
25 Jul 2024 · Paper summary
Paper summary: OpenHands: An Open Platform for AI Software Developers as Generalist Agents
OpenHands (formerly OpenDevin) is an open platform for agents that work like software developers, writing code, using a shell and browsing, with sandboxed execution and built-in benchmarks.
23 Nov 2023 · Paper summary
Paper summary: GAIA: a benchmark for General AI Assistants
GAIA is a benchmark of 466 real-world questions that are easy for people but need reasoning, browsing, multimodality and tool use; humans scored 92% against 15% for GPT-4 with plugins.
4 Aug 2023 · Paper summary
Paper summary: Learning to Model the World with Language
Dynalang treats language as a signal for predicting the future: an agent learns a multimodal world model and acts from imagined rollouts.