Skip to main content
ZICQ

Topics Agent Frameworks

Agent Frameworks

AI Agent frameworks compared: LangGraph, AutoGen, CrewAI, Letta — design philosophy, tool use model, and production fit.

· Aug 15, 2026 #Agent#agent#LangGraph#LangChain#AutoGen#CrewAI#Letta · 30 / 4318 articles

News · 2026-11-19

GUARD Stabilizes Recursive Self-Improvement with Directional Regret

GUARD targets recursive self-improvement in agents that keep a frozen base model but continuously edit an external harness containing prompts, skills, tool rules, and memory. It converts tasks into pre-execution self-checks, uses directional regret to distinguish newly broken behaviors from unresolved ones, ranks parallel edit candidates through critic-free group-relative fitness, and accepts an update only when a Soft-HURT gate finds that its gains outweigh its regressions. The paper reports g

News · 2026-11-18

AnyS1 Uses a 4B Pointer Model to Cut Routine Agent Decision Costs by up to 91%

AnyS1 addresses a structural inefficiency in LLM agents: invoking a large System-2 model at every workflow step, even when the next action must be selected from a small set of known options. The framework mines more than 1,000 public procedural skills for grounded, enumerable decision points and uses execution-grounded trajectories collected in isolated containers to construct Any-DP. At runtime, a calibrated 4B pointer model resolves routine choices in one forward pass and invokes System-2 onl

News · 2026-11-18

PIPETTE: Diverse Red-Teaming and On-Policy Training for Robust LLM Agents

PIPETTE targets a critical mismatch in indirect prompt injection defense: red-teaming systems can achieve high attack success rates (ASR) while repeatedly discovering attack families that add little value to adversarial training. Building on evidence of mode collapse in RL-Hammer and PISmith, the framework combines a dictionary-learning-based reward that rewards broader strategy coverage with an on-policy adversarial-training loop tailored to IPI. The paper reports near-zero ASR on static attac

News · 2026-11-18

Game2Player Co-Develops Game Environments and Players to Scale Visual-Agent Training

Game2Player frames gameplay-data collection as a closed-loop co-development process rather than a one-off annotation task. A Builder agent instruments existing web games, exposes internal state, and defines tasks with state-based progress measures and success checks. Coding and VLM players then test those tasks, report failures, and trigger repairs, so each play both validates the environment and produces training trajectories. The team reports 568 verified web games, 3,436 tasks, 15.7K traject

News · 2026-11-18

IDEA Audits GUI-Agent Visual Backdoors Step by Step Without Trigger Knowledge

IDEA addresses a supply-chain failure mode in open-weight GUI agents: a compromised checkpoint passes clean acceptance tests but changes behavior only when a visual trigger appears. Rather than matching a known trigger pattern, it audits an intent-dominance inversion at every agent step. Intent Counterfactual Probing measures how strongly a redirected instruction moves the decision state, while Region Causal Audit measures how strongly ablating a candidate screen region moves it. A gated confor

Info · 2026-11-18

Google DeepMind Releases Vibe Robotics, an Antigravity Plugin for Real Robots and MuJoCo Simulation

Google DeepMind's GDM Robotics team has published Vibe Robotics, a plugin for Antigravity agents that packages robotics engineering practices as reusable skills. It is intended to work with physical manipulators and MuJoCo simulations; the team reports testing on SO-101, YAM, and Franka Duo. The project connects agent-level planning with motion planning, visual servoing, and collision avoidance, with claimed reductions in tokens and trial-and-error during robot bring-up. The public material con

News · 2026-11-18

Airbnb Runs Coding Agents on Amazon Bedrock with Task-Level Cost Control

Airbnb has published the operating model behind AirChat, its internal coding-agent platform, where requests pass through an LLM gateway to Amazon Bedrock. Before rollout, the company scores models on representative engineering tasks at each reasoning effort; afterward, it reconciles task-level evaluations with CloudWatch latency, throttling, token-mix, and spending telemetry. The relevant decision metric is not the price of one API call, but the list-price cost of every token consumed across all

News · 2026-11-18

Snapchat Lets Claude Code and Cursor Build AR Lenses from Natural-Language Prompts

Snapchat now allows developers to create AR Lenses with AI coding assistants such as Claude Code and Cursor. A creator can describe the desired experience in natural language, while the assistant is expected to carry the process through implementation. The important shift is not merely a conversational interface: it turns intent translation, code and asset generation, build execution, and debugging into a potentially agentic authoring workflow. The available report does not disclose the integra

Info · 2026-11-18

Mastra Factory Open-Sources an Agentic Software Factory Built for Slack

Mastra has open-sourced Mastra Factory, the agentic software-production system its team uses to maintain the TypeScript agent framework. Rather than a conventional IDE, it exposes a configurable, @-mentionable kanban in Slack and orchestrates triage, planning, coding, review, and delivery through durable agents, conversational memory, a cited Markdown wiki, PR-comment signals, and a customizable cloud sandbox. The team reports that, three months after adoption, most sessions are launched through

News · 2026-11-18

Stretto Rebuilds Bankruptcy Intake with Amazon Nova 2 Sonic

Stretto, the AWS Generative AI Innovation Center, and CI&T deployed Tara, a voice agent for multi-stage bankruptcy intake. Unlike a conventional ASR-to-LLM-to-TTS pipeline, Tara uses Amazon Nova 2 Sonic through Amazon Bedrock as a native speech-to-speech model, while Twilio streams caller audio to an ECS/Fargate voice bridge. Claude 3 Haiku extracts typed fields asynchronously after the call, while deterministic attorney-defined rules—not the model—determine lead qualification. The most reusabl

News · 2026-11-18

AWS Publishes a Hub-and-Spoke Pattern for Inspecting AgentCore Gateway Traffic to Private APIs

AWS has published and tested a reference architecture for connecting Amazon Bedrock AgentCore Gateway to private APIs and MCP servers in a centrally governed hub-and-spoke network. The network account owns a VPC Lattice Resource Gateway, routes its traffic through an inspection VPC and Transit Gateway, and shares backend Resource Configurations with the AgentCore account through AWS RAM. The repository includes CloudFormation templates, test scripts, troubleshooting guidance, and a 14-check val

Info · 2026-11-18

Microsoft Agent Framework PyPI Package Compromised by Credential-Stealing Malware

Microsoft's `agent-framework` package on PyPI was compromised with a credential-stealing payload, according to the supplied headline from Socket.dev. The potential blast radius extends beyond a developer's workstation: Python packages execute code through build, installation and import paths, while CI runners often hold publishing tokens, Git credentials, cloud IAM material and model-service keys. A malicious release can therefore turn routine dependency installation into secret exfiltration and

News · 2026-11-18

DeltaHarness: Self-Evolving Tool-Execution Policies for Resource-Adaptive Agents

DeltaHarness treats the agent execution harness as an adaptive control layer rather than a fixed script. Starting from cold-start measurements, its QHI loop uses observed responses to update runtime dispatch, revise physical execution plans, and select follow-up experiments; an independent confirmation panel gates policy replacement. The abstract reports a 30.0% reduction in complete-delivery latency for a revised materialization policy under fixed dispatch on one video-editing task, evaluated

News · 2026-11-17

More Reasoning Raised Confidence, Not Accuracy, in LLM Shopping Agents

A study of multimodal shopping agents evaluates ten frontier models on three product-choice benchmarks derived from WebShop, Amazon SQID, and Wayfair WANDS. Raising reasoning effort produced a statistically significant increase in stated confidence, but no statistically meaningful improvement in decision quality on any benchmark. Most models remained poorly calibrated at every tested effort level. Errors were also correlated: when several models failed the same task, they often selected the same

News · 2026-11-17

Resolution Hypothesis: Noisy Evaluation Bottlenecks Agent Evolution

The paper reframes agent evolution as a measurement problem rather than primarily a search problem. Across seven evolution methods, six models, and more than 10 million LLM calls, it finds that helpful prompts, skills, and memories often exist, yet the selector is no better than random at distinguishing them from harmful candidates. At the method level, typical performance gaps of only a few percentage points are smaller than the variation between independent runs of the same method, making man

News · 2026-11-17

CADMAGEDDON: Open-Source Local Agent Reconstructs Editable CAD from Meshes

CADMAGEDDON frames mesh-to-CAD reconstruction as a sequence of visual decisions: infer what to build, invoke geometric construction tools, inspect the result, and correct remaining errors. FreeCAD integration keeps outputs editable and compatible with existing workflows, while the open implementation targets single-GPU and current-generation laptop deployment for proprietary geometry. The paper reports that a system using the mid-sized Qwen3.8-27B model outperforms prior approaches—including sy

News · 2026-11-17

RepoFactory Evolves Verifiable Coding Environments, Boosting RepoHorizon by 22.6 Points

RepoFactory converts real software repositories into 10,159 verifiable training environments spanning 17 domains. Agents receive a project specification and must build a complete implementation from scratch, while an Environment Evolution Loop uses proxy feedback to optimize validity, structural coupling, and implementation scope. Training Qwen3.5-35B-A3B on these environments produces RepoHorizon-35B-A3B. The paper reports a 22.6-percentage-point gain on NL2Repo-Bench and an 8.5-point average

News · 2026-11-17

LLM Agents Design Their Own Self-Improvement Algorithms

The paper asks whether an LLM agent still needs an externally specified self-improvement loop. Given a frozen model, a general objective, a compute budget, and tools for collecting experience and editing text artifacts, the agent is allowed to invent a context-based self-learning procedure in a gameplay environment. The abstract reports that sufficiently capable models improve under this setup, using patterns resembling self-regulated learning and self-generated curricula, and that the resulting

News · 2026-11-17

OpenAI Launches ChatGPT for Sales and Expands 30+ Partner Plugins

OpenAI has introduced a Sales plugin in ChatGPT alongside 30+ new or expanded partner plugins. The announced workflow covers prospect research, pitch preparation, CRM updates, and progressing deals, positioning ChatGPT as an action surface across external business systems rather than merely a content generator. The release note, however, does not disclose the underlying plugin runtime, supported models and regions, permission model, pricing, or measurable reliability and cost. Production evalua

News · 2026-11-17

Cutting Token Use by Up to 71% and Latency by 80% for Large JSON Workloads on Bedrock AgentCore

AWS documents a reusable context-engineering pattern for a Flightscape airline-operations agent running on Amazon Bedrock AgentCore with more than 40 tools and JSON payloads above 20 MB. The design separates live graph state from persisted state, stores heavy artifacts in Amazon S3 under deterministic SHA-256 content addresses, pages oversized lists through a result handle, and replaces large nested dictionaries with compact structural descriptors. The model receives a summary first and retrieve

News · 2026-11-17

AWS FDE Expands to 15 Countries, Taking Enterprise Agents to Production in 45-Day Sprints

AWS has expanded Forward Deployed Engineering, or FDE, to 15 countries for commercial and public-sector customers. The engagement model uses fixed-price 45-day sprints, with AWS engineers, scientists, product managers, customer engineers, and specialized AI agents working together inside the customer’s own AWS account. FDE is a production-delivery framework rather than a new foundation model. Its architecture combines a reusable business context layer, agent-driven workflow implementation, the

News · 2026-11-17

Tasks in Gemini Is Free in Six Countries and Gains Deeper Workspace Integration

Google says Tasks in Gemini is now free in six countries, with additional features beginning to roll out in the United States. The central direction is deeper integration with Google Workspace, allowing Gemini to act on users’ behalf when appropriately authorized. The available announcement does not identify the six countries, enumerate the new features, specify supported Workspace services, or disclose the underlying model, orchestration design, confirmation policy, and measured success rates.

News · 2026-11-17

LoanPro Deploys Multi-Agent Loan Servicing Assistant, Cutting Call Time by Up to 15%

LoanPro, working with the AWS Generative AI Innovation Center and AllCloud, has deployed a multi-agent assistant for mortgage servicing teams on Amazon Bedrock. Built with the Strands Agents SDK, the system separates account retrieval, account summarization, next-best-action recommendation, and response composition into specialist agents. Amazon Bedrock AgentCore Runtime and Gateway connect those agents to LoanPro’s production APIs through the Model Context Protocol, while API Gateway, Lambda, a

News · 2026-11-17

AWS Quick Sessions at re:Invent 2026: Zero-Trust Governance, Multi-Agent Workflows and MCP

AWS has published a full guide to its AWS Quick sessions at re:Invent 2026, scheduled for November 30 through December 4. Rather than announcing a new foundation model, the program lays out how Quick is intended to operate as an enterprise AI workspace: shared dashboards, agents and automations; governed connectors; zero-trust identity; GraphRAG and knowledge graphs; persistent memory; MCP-based actions; and multi-agent workflows. The sessions range from tenant isolation and compliance to conve

News · 2026-11-17

AWS Introduces a Three-Phase Responsible AI Roadmap for Agentic Systems

AWS has published a three-phase responsible-AI maturity framework for platform and application teams building agentic systems. Rather than introducing a new model, it connects model admission, risk classification, least-privilege tool access, human approval, continuous evaluation, OpenTelemetry tracing, and CI/CD release gates to an AgentOps feedback loop. The accompanying GitHub assessment skill scores readiness across eight responsible-AI dimensions on a 1–5 scale. The guidance is particularl

News · 2026-11-17

NVIDIA Unveils a Scientific Agent Stack and Rubin HPC at SC26

NVIDIA introduced Nemotron for Science at SC26, combining the Nemotron 3.5 Super VL open model with a chemistry dataset and post-training recipes for life-science and chemistry workflows. The stack can run on DGX Spark, workstations, or existing supercomputers. NVIDIA also previewed the Vera Rubin platform, with expanded Vera CPUs for agent tool execution and Rubin HPC GPUs spanning FP64 to FP4, sparse acceleration, and seventh-generation NVLink. NVIDIA also expanded its CUDA-Q work with IBM an

Info · 2026-11-17

FrontierHarness Eval v1.1: 64% of Coding-Agent Inference Spend Went to Failed Runs

FrontierHarness Eval v1.1 reports a cost attribution across 27,280 coding-agent runs: 64% of inference spend was consumed by runs that ultimately failed. The metric is a spend share, not a 64% failure rate, and should not be read as one. It instead identifies a costly tail in which failed trajectories persist through long contexts, repeated tool calls, and retries. The practical metric is cost per uniquely solved task, accompanied by failure taxonomy and quality-cost Pareto analysis. Because th

News · 2026-11-17

A Preference-Grounded Framework for Evaluating LLM Ambiguity Resolution

The paper introduces a preference-grounded simulated-user framework for controlled, multi-turn evaluation of ambiguity resolution. Rather than allowing an unconstrained simulator to improvise, the simulator responds only within explicit user disambiguation preferences, reducing evaluator-side hallucination and over-disclosure. The framework tracks three observable dissatisfaction signals—incorrect assumptions, repeated clarifications, and irrelevant clarifications—so task accuracy is no longer t

News · 2026-11-17

SPLICE Turns Agent Prompt Optimization into Constrained Combinatorial Search

SPLICE reframes natural language optimization for AI agents as constrained combinatorial search over block-structured, selectively mutable policy documents. Contract blocks such as safety clauses, tool schemas, and output formats can remain verbatim, while mutable sections are optimized through local textual critique. Elitist beam search, diversity slots, and a ratio-based length regularizer target validation regression, lineage collapse, and uncontrolled prompt growth, with formal bounds stated

News · 2026-11-17

VEX Trains Computer-Use Agents from Audio-Visual Tutorials and Replays Their Own Actions

VEX reframes video demonstrations as audio-visual experience rather than screenshot-only action traces. In a watch-act-rewatch loop, a 9B video-SALMONN-VEX agent extracts procedural guidance from tutorial narration and imagery, acts on a GUI, and replays its own execution to check progress. The paper also introduces AGGO, a reinforcement-learning method that optimizes video-derived guidance for downstream action prediction, and adds more than 24,000 real-world tutorial videos to AgentNet, OSWorl