The Hub for AI Coding Models, Tools & Workflows

Strands Agents (harness-sdk) python/v1.57.1 (2026-09-25) adds bidi snapshot capture/restore, a vended a2a_client tool, Agent.shutdown() for scope cleanup, CLI Quickstart/Customize/Export replacing Q&A setup plus /tools, and MCP TS client 2.0. PyPI strands-agents 1.57.1 is live.

Google ADK Python v2.10.0 (GitHub 2026-09-25; changelog 2026-09-24) adds experimental skill lifecycle controls (EPHEMERAL, active-skill caps, unload_skill via ADK_ENABLE_SKILL_LIFECYCLE=1), a MongoDB toolset with vector/hybrid search, eval efficiency metrics for duration/tokens/model calls, and better OpenAI reasoning param adaptation plus reasoning-token reporting. PyPI google-adk 2.10.0 is live.

Comet Opik 2.2.81 (2026-09-25) routes scored entities into annotation queues, optionally provisions an extended free-form SQL ClickHouse user, registers 19 canonical providers so model prices load, and fixes Gemini usage, streamed Mistral reasoning, and slots=True dataclass encoding in the SDK.

OpenHands Agent Canvas v1.24.0 (2026-09-25) adds header workspace-folder toggles, read-only shared automation conversations on cloud, MCP OAuth credentials preserved on cloud saves (skip consent when tokens still work), OpenAI subscription caches scoped by backend, and Node.js ≥24.

Langfuse v4.46.0 (2026-09-25) ships basic skill management (PR #17803) via unstable REST (list/create/get/patch/delete skills, versions, file content), streamlined experiment side-by-side comparison, formatted JSON on dataset items, optional 100-row tracing page size, assistant respect for data-retention, and v3 dataset-run-items reads on replicas.

Cline Desktop v0.0.36 (2026-09-25) skips system HTTP(S) proxies for local hub checks, fits oversized windows to small/scaled displays, loads plugin slash commands on workspace open, and runs Bedrock GPT-6/GPT-5.6 without cross-region inference (India via in. profiles).

DeepSeek AI (@deepseek_ai) has officially launched the standalone desktop application for DeepSeek Harness (dsh) across macOS and Windows. Designed to eliminate the engineering overhead of managing CLI background daemons and local sandbox sockets, DeepSeek Harness Desktop integrates the web UI, the Cordis plugin execution engine, and isolated sandboxing into a native local client. It delivers out-of-the-box support for visual trajectory playback, arbitrary session branching, and multi-model sub-agent delegations.

Pydantic AI 2.50.0 (2026-09-25) adds DecisionModel (TypeSafeModel implements it), named route questions with optional decision_route_threshold, Gemini 3.8 Live realtime, durable-context hooks, and audio_seconds usage pricing.

DSPy 3.4.0 (2026-09-25) integrates Jev through TypeSafe with experimental Noul/Choice/Score decision types and ReAnchor calibration; the LM layer prefers native lm15 engines, adds LocalInterpreter and async ReActV2. 3.4 is the LM transition release; 3.5 is the migration deadline.

Local LLM inference runtime Ollama (@ollama) officially rolled out v0.40.0. The standout architectural milestone is the default adoption of Apple's native MLX framework on all Apple Silicon Mac devices (M1 through M5), replacing the legacy CPU/Metal compute paths. Supporting major architectures like Qwen 3.8 and Llama 3 natively, the update harnesses Apple's unified memory bandwidth, slashing time-to-first-token by 42% and nearly doubling continuous generation throughput without manual configuration.

OpenAI Codex CLI rust-v0.157.0 (2026-09-25) adds GPT-6 Sol/Luna (incl. Amazon Bedrock), enables fullscreen transcripts and auto background-server startup by default, adds an f shortcut to fork locked conversations, and enables /import in remote and local daemon sessions.

Researchers from Johns Hopkins University have introduced Knowledge Pull Requests (KPRs, arXiv: 2609.26634), an interpretable framework for continual document authoring. Treating knowledge curation analogously to software pull requests, KPRs decouple factual claim extraction from textual diffs and automate contradiction detection, outperforming blind text regeneration in enterprise RAG systems.

Tencent and Zhejiang University's REAL Lab have open-sourced IterSynth (arXiv: 2609.29444, GitHub: Tencent/IterSynth), a groundbreaking paradigm for deep-search AI agents. Breaking free from the rigid single-agent ReAct framework where planning, evidence retrieval, and synthesis are tightly coupled—which typically causes severe context saturation and attention degradation—IterSynth alternates between a specialized Planner and a Synthesizer governed by an evolving summary state. Powered by Role-Decoupled Policy Optimization (RDPO), IterSynth-8B hits an average score of 50.7 across long-horizon benchmarks like BrowseComp and Xbench-DS, outperforming all prior <=8B search agents by +4.2%.

Security and systems researchers have introduced AgentKernel (arXiv: 2609.29647), a trust-native operating system designed specifically for autonomous AI agents. As modern agents routinely ingest untrusted external data and execute privileged tools, traditional application-level guardrails fail to prevent prompt injections, memory poisoning, and tool abuse. AgentKernel establishes a mandatory, non-bypassable OS substrate structured around four architectural pillars—Identity, Perception, Cognition, and Execution—bringing deterministic capability confinement and information flow control to the semantic plane.

A robotics research consortium led by Tom Silver and Matteo Merler has open-sourced groundbreaking findings titled 'Coding Agents for Generalized Task and Motion Planning Problems' (arXiv: 2609.30233, GitHub: tomsilver/robocode). Overcoming the fragility of human-engineered heuristics in Task and Motion Planning (TAMP), the authors deploy autonomous coding agents (Claude Code Opus 5, Codex) to synthesize generalized symbolic-geometric programs through closed-loop simulator interaction. Across 98,000 evaluation episodes spanning 28 benchmark environments, the synthesized programs achieved 56% to 95% success rates, outperforming hand-crafted planners while using an order of magnitude less compute.

A research collective including Yilun Du, Lvmin Zhang (creator of ControlNet), and Haotian Zhang has released 'Training Object Permanence in World Models' (arXiv: 2609.28654, Project: object-permanence.world). While video generative models produce photorealistic scenes, they routinely violate physical object permanence—causing occluded objects to magically vanish, morph, or teleport. The researchers introduce WROP, a cognitive benchmark spanning 150 tasks and 1.5 million training trajectories, alongside PWM-WROP, an open-source 16B world model trained on AWS Trainium2 that secured top rank among video continuation models in blind Elo evaluations.

Mastra shipped @mastra/core 1.71.0 on 2026-09-24: agent.stream starts each tool as soon as its args complete (eagerToolExecution on by default), cutting idle seconds on multi-tool steps; adds planTraceAggregate(), observability getFeatures discovery/feedback flags, and RE2-backed dataset schema regexes to close a DoS vector.

Researchers from Zhejiang University's OmniAI Lab have open-sourced Spatial-Interactor alongside the LSI-108K interactive spatial reasoning dataset (arXiv: 2609.23038). Addressing severe 3D hallucination in passive vision-language models, Spatial-Interactor introduces closed-loop active perception where an agent's physical camera-control actions inform subsequent spatial inferences, fully accessible under the Apache 2.0 license.

AG2 v1.1.0 (2026-09-24) adds TypeSafe Jev decision-only agents via TypeSafeConfig + response_schema, and strips reserved ag:/a2a: context keys on A2A/AG-UI/A2UI/NLIP to block remote pre-approval of gated tools.

Researchers have published a seminal framework titled 'Self-Organizing Agent Teams Learn to Reason Together' (arXiv: 2609.22682). Overcoming the rigidity of hardcoded hierarchical agent structures, the architecture leverages graph rewriting to allow LLM agents to dynamically self-organize sub-teams and communication topologies, improving complex problem-solving rates by 27.4%.