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 / 4348 articles

News · 2026-11-06

SIRIS Shows LLMs Often Misjudge Simulator Error

SIRIS evaluates whether LLM agents can distinguish a harmless simulator distortion from one that invalidates a systems-design recommendation. Its paired workloads preserve the candidate designs, simulator limitation, and original recommendation while applying perturbations that change predicted behavior; a reference simulator then determines whether the recommendation remains valid. Across 13 frontier LLMs, recognizing a limitation did not reliably lead to the right design decision, and even th

News · 2026-11-06

CPU-ARC Compiles LLM Reasoning Traces into Reusable CPU Programs

CPU-ARC introduces agentic reasoning caching, a method that compiles LLM reasoning traces into CPU-executable agentic reasoning caches. By separating problem solving from solution execution, the resulting programs can reuse prior reasoning for new instances of related tasks without repeatedly invoking a large language model. The paper reports performance comparable to frontier LLMs across difficult mathematical and scientific tasks, along with explainability, continued trainability, and robustne

News · 2026-11-06

ADB Reveals Authorization Drift in LLM Agents

The Authorization Drift Benchmark (ADB) formalizes how LLM agents exceed authorized boundaries across data, tools, scope, and time. It contains 648 paired tasks across seven domains and validates both compliant paths and attractive unauthorized shortcuts through real agent rollouts rather than model self-critique. Across 13 LLMs, three agent frameworks, and 25,207 trajectories, stronger models tended to exhibit higher drift, with frontier-capable models reaching rates of 0.43–0.61. However, 86.

Info · 2026-11-06

XingChen-AGI Open-Sources Xing4.0, a 29B Sparse MoE with 256K Native Context

XingChen-AGI, part of China Telecom Artificial Intelligence Technology, has open-sourced Xing4.0-29B-A4B, a sparse mixture-of-experts model with 29 billion total parameters and 4 billion active parameters per token. It uses 64 routed experts with four selected for each token, plus one shared expert, and provides a native 256K-token context window that can be extended to 512K. The model combines MLA, mHC, and MTP to target agent planning, tool use, and complex engineering workflows. Its training

Info · 2026-11-06

Zalando Open-Sources AIB for Scoped OAuth Delegation to AI Agents

Zalando Incubator has open-sourced the Agentic Identity Broker (AIB) under the MIT License. AIB gives AI agents scoped and revocable access to MCP services without exposing the user's underlying OAuth tokens. Users grant consent per agent and service, while the broker stores third-party tokens in encrypted form and performs RFC 8693 token exchange at the gateway. The v0.1 release is reportedly running in production, connecting hundreds of agents to dozens of internal and external MCP servers. Pl

News · 2026-11-06

Closed-Loop Agent Evaluation: Layered Measurement and Self-Calibration

This paper proposes closed-loop evaluation for deployed AI agents, treating evaluation as performance estimation under biased raters and censored outcomes. It analyzes agents across span, trace, conversation, and engagement levels, measuring execution, substance, trajectory, and results while enforcing conduct as a cross-cutting constraint. The framework combines provenance, adjudication, versioned rules, and two feedback loops: an inner loop audits evidence, rules, and metrics, while an outer

News · 2026-11-06

A Theoretical Framework for Self-Directed AI Evolution—and Its Alignment Trap

The paper develops a theoretical framework for comparing Darwinian, Lamarckian, and agent-directed evolution after an environmental change. Learning and broader behavioral search can accelerate adaptation, but search effort is not uniformly beneficial: its optimum depends on learning quality, task difficulty, and experimentation cost. The central safety result is an alignment trap. When the evaluation objective is misspecified, increasing search accelerates convergence toward the wrong behavior

News · 2026-11-06

TRAP Dynamically Adjusts Agent Authority, but Gradual Degradation Exposes a Blind Spot

The paper introduces the Trust Responsive Authorization Policy (TRAP), which converts runtime trust, operational stability, and contextual risk into an Authorization Readiness Score and maps that score to discrete authority levels. Its asymmetric control logic immediately revokes authority after a sharp readiness drop, while restoration requires consecutive observations above a band edge and margin. Across 100 matched Monte Carlo traces, TRAP reacted promptly and stably to abrupt degradation. Ho

News · 2026-11-06

Two-Ledger Model Attributes Multi-Agent AI Costs to Business Transactions

Enterprise AI billing is typically aggregated at the SKU and billing-period level, making it difficult to determine what an individual order, claim, or case actually cost across models and agents. The paper proposes a Two-Ledger Model that keeps a FOCUS-conformant settlement ledger and joins it with an invocation-grain attribution ledger built from OpenTelemetry GenAI telemetry. Trace context, A2A/MCP metadata, business-key correlation, and probabilistic matching form the correlation spine. A v

News · 2026-11-06

KDASO Dynamically Optimizes Skill Libraries for Liability Adjudication

KDASO is an automated skill-optimization framework for liability adjudication. It combines domain knowledge, recent case data, and LLM execution traces to continuously induce, refine, and route interpretable skills. Neighborhood-conditioned induction helps the skill library track evolving scenarios, trajectory-based refinement reduces the semantic gap between written rules and model behavior, and context-budgeted routing prevents critical skills from being omitted as the library grows. The pape

News · 2026-11-06

Octopus Rebuilt Multi-Agent Reinforcement Learning Around a Shared KV DAG

Octopus is a research framework for long-context multi-agent reinforcement learning that replaces duplicated context branches and text-summary communication with a GPU-resident directed acyclic graph of KV states sharing one model backbone. Its Zero-Copy Radix Forking creates branches as O(1) cursor views over shared physical KV pages, while Attention-Only KV Mounting directly exposes sibling terminal KV blocks to full-attention layers without propagating incompatible recurrent states. The fram

News · 2026-11-05

IVG-Bench Tests Vision-Language Models in Closed-Loop Visual Interaction

Interactive Visual Gym evaluates vision-language models as real-time assistants for physical tasks by jointly simulating executable environments and human users. The resulting IVG-Bench contains 315 scenarios spanning cooking, household manipulation, and mobile interfaces, with explicit treatment of response latency, factual grounding, actionable guidance, execution errors, and out-of-plan user behavior. The best evaluated model, Gemini 4 Argon, completed only 53.4% of tasks within the time limi

News · 2026-11-05

FINCOSMOS Exposes Unreliable Financial Agents: None of 150 Sessions Passed

FINCOSMOS introduces an open-world evaluation paradigm for financial agents operating on live Web data. Agents must discover data, reconstruct a user's hidden strategy, build point-in-time-correct backtests, and deliver independently verifiable results. Grading relies on executable checks rather than a reference answer: the system recomputes reported metrics, validates the hidden strategy specification, and rejects runs that fail any of 15 veto checks. In 150 sessions involving four coding agen

News · 2026-11-05

ReproBridge Bakes Reproduction Knowledge into ML Artifacts

ReproBridge introduces an agent workflow that converts papers and raw repositories into standardized reproduction artifacts: a refined codebase, environment setup, experiment launchers, and a reproduction guide. An independent auditor validates execution evidence. On ReproBridge-Bench, covering 1,029 tasks from 387 experiments across 36 papers, the refined artifacts raise mean reproduction success across four models from 76.4% to 92.6%, while downstream API spending falls by 19.7%. The spread b

News · 2026-11-05

PCBRouteAgent Brings Hierarchical Planning to Complex PCB Routing

PCBRouteAgent is a multimodal agent designed to automate printed circuit board routing by separating board-level strategy from geometric execution. Its Global Routing Planner combines PCB images with structured electrical and geometric data to specify connectivity topology, routing order, corridors, and layer transitions. A Detailed Routing Executor then applies these plans through KiCad's native routing engine, performs bounded local recovery, and returns structured feedback for targeted revisi

News · 2026-11-05

HEELO Packages Agent Capabilities into Verifiable Primitives for Real-World Tasks

The paper introduces Agent-Native Primitives and HEELO, a harness-engineering framework that turns resources such as tools, terminals, development environments, and navigation interfaces into reusable execution units equipped with resident LLMs. HEELO composes these units into task-specific graphs, verifies outcomes against environmental evidence, and revises or admits primitives only when failures expose a concrete gap or a candidate passes validation. Across more than 20 benchmarks, the repor

News · 2026-11-05

Agent Evolution Is Not Monotonic: EvoAudit Exposes the Engineering Limits of Recursive Improvement

This study evaluates four methods for evolving agent harnesses across general and coding tasks, focusing on whether benchmark gains translate into reliable, efficient, and transferable improvements. All 16 evolved agents solved previously uncovered tasks, but they also lost tasks they had previously handled. Reliability gains often increased deployment cost, and the highest-scoring agent did not produce the largest average improvement across the four evaluation pools. The paper argues for preser

News · 2026-11-05

LangChain Adds Managed Browser and Human Login Handoff to Deep Agents

LangChain has added a managed browser to Managed Deep Agents, allowing agents to interact with websites in a controlled execution environment while handing sensitive steps such as authentication back to a human. The browser tools are designed to keep context usage under control, reducing the burden of feeding large amounts of page state into the model. The update brings web navigation into the managed agent execution path, but the announcement does not specify tool APIs, sandbox isolation, brow

News · 2026-11-05

Perplexity Launches Always-On Teammates for Multi-Agent Work

Perplexity has introduced Teammates, an always-on, work-focused agent system that lets users assign different roles to multiple agents within a shared workspace. Built on top of Perplexity Computer, which can deploy agents on a device to execute tasks, Teammates is designed to support parallel, ongoing, large-scale work while reusing existing projects, skills, and memory from the user's Perplexity tools. The release puts Perplexity in direct competition with products such as OpenAI's Dots and A

News · 2026-11-05

AWS Launches Web Answer for Cited Web-Grounded Responses

AWS has announced the general availability of Web Answer on Amazon Bedrock AgentCore. This managed connector accepts a natural-language question, plans and executes parallel searches, reads key pages to verify facts, and searches again when the evidence is insufficient. It then returns a synthesized answer whose claims are linked to source URLs through a single Model Context Protocol tool call. The service packages search, page retrieval, synthesis, and citation attribution into an AWS-operated

News · 2026-11-05

LangChain Adds Async Subagents to Keep Long-Running Tasks Off the Critical Path

LangChain has introduced async subagents in Deep Agents, allowing a main agent to delegate long-running work to background workers while keeping the conversation responsive. An orchestrator coordinates task execution, notifications, and human approvals, separating the interactive path from slow tool or compute paths. The design can improve responsiveness in research, code execution, and other multi-step workflows, but production adoption still requires careful handling of cancellation, timeouts,

News · 2026-11-05

Meta Adds Real-Time Voice Mode and Animated Avatars to Muse AI

Meta has begun rolling out real-time voice conversations for the Muse AI agent. On iOS, its animated avatar responds in sync with spoken output, while users can request a desired vocal tone in natural language or select presets from avatar settings. The Verge reported generally quick responses, with occasional pauses while the agent looked up information. The update advances Muse toward multimodal, low-friction interaction, although Meta has not disclosed its voice architecture, latency figures,

News · 2026-11-05

Claude Connects to Bank Accounts for Read-Only Personal Finance Analysis

Anthropic has introduced Personal Finance for Claude, using Plaid to connect bank and credit-card accounts through a read-only integration. Claude can summarize spending by category, estimate bills due before the next payday, assess whether a major purchase is affordable, and estimate potential monthly savings. The feature turns transaction data into conversational financial analysis, but its usefulness will depend on institution coverage, transaction categorization quality, refresh latency, an

News · 2026-11-05

LangSmith BYOC Expands to AWS, Azure, and Google Cloud

LangChain has announced that LangSmith BYOC is generally available on AWS and entering beta on Microsoft Azure and Google Cloud. LangChain manages the platform, while traces, evaluation data, and agent deployments remain inside the customer's cloud account. The release targets enterprises that need centralized AI observability without surrendering control of their data plane. Its practical value lies in deployment topology and governance rather than new model capabilities; buyers still need to v

Info · 2026-11-05

Rivet Open-Sources an Actor-Based Agent Orchestrator Backed by Tiered S3 Storage

Rivet has open-sourced an actor-based orchestrator designed for long-running agentic workloads. Instead of provisioning a full sandbox for every agent, it keeps each agent in a lightweight actor with in-memory state, durable SQLite storage, WebSocket support, and idle hibernation. The persistence layer combines a replicated SSD write tier with background batching to S3. Rivet reports 72 KB baseline actor memory, 4.9 ms p99 startup latency, 1.07 ms p99 durable writes across three availability zo

News · 2026-11-05

Microsoft Releases PDD Kit to Formally Verify Agent-Generated Code

Microsoft Research has introduced PDD Kit, a research prototype that brings Proof-Driven Development to agentic software development. Developers approve a plain-language intent artifact, after which the system derives Verus specifications, proofs, and tests for trusted boundaries. Verus uses Z3 to deterministically check whether the implementation satisfies the approved contract for every input covered by the model. The toolkit's Pet Store case study applies the workflow to Rust login logic, ca

News · 2026-11-05

NVIDIA Connects Frontier AI Agents to Omniverse Simulation Workflows

NVIDIA is showcasing workflows that combine frontier AI models with Omniverse libraries to create, inspect, and modify simulation environments through natural-language instructions. The demonstrated applications span editable 3D assets, warehouse digital twins, robot learning, autonomous-driving tests, sensor validation, and GPU-accelerated physics and rendering. The strongest disclosed engineering result is a warehouse scene with 80 conveyors, 40 robot arms, and 800 boxes that maintained 30 FP

Info · 2026-11-05

DialFork Benchmarks Procedural Hallucination in Conversational Agents

Idiap Research Institute and Logitech have introduced DialFork, an open benchmark for measuring whether conversational agents remain faithful to an assigned workflow even when their claims remain fluent, factual, and their tool calls are correct. The benchmark isolates procedural hallucination: omissions, reordering, or replacement of required conversational steps such as asking, confirming, and reporting results. DialFork generates more than 36,000 dialogues by replaying correct workflows and

News · 2026-11-05

SlackBench Benchmarks Collaborative Agents Across Code and Workplace Conversations

SlackBench introduces a reproducible environment for evaluating agents on collaborative research and engineering workflows. It links multi-party workplace conversations with evolving GitHub changes and Google Docs decisions, covering heterogeneous-source search, abstention, goal-directed summarization, and circumstance inference across long conversations. Across 33 frontier models and agent harnesses, the strongest system reaches roughly 80%, while most remain below 66%. The results show that r

News · 2026-11-05

ReScopeBench Evaluates the Experience Lifecycle of Long-Lived Agents

ReScopeBench evaluates how long-lived agents acquire, reuse, revise, and lose experience as environments evolve. It organizes task streams around shared experience across 18 interactive environments, including silent changes that invalidate previously learned information, and evaluates 20 models while tracking their saved experience over time. The results separate two persistent bottlenecks. Models benefit from retaining experience, but they struggle to update stale knowledge when the environme