Mastra launched the Custom Harness Program, in which forward-deployed engineers embed with customer teams to co-build tailored agent harnesses. The program includes observability, evals, custom MCP, human-in-the-loop workflows, auth and RBAC, plus early access to the Mastra Enterprise Platform. Mastra cites Workday, which moved off Amazon Bedrock Agents in 31 days, cut 3,000 lines of code and lowered vector-storage costs by 98%.
Buoyant announced Kinference, an Apache v2 open source control plane that routes OpenAI- and Anthropic-compatible inference traffic to vLLM workers on Kubernetes, adding token rate limiting, per-tenant usage and cost tracking, KV-cache-aware load balancing, and YAML inference policies. In the project's public benchmarks, Kinference cut median time-to-first-token by 95% and median end-to-end latency by 45% versus standard Kubernetes service load balancing on one workload, and served 7.4x the goodput on shared-prefix and about 3.5x on multiturn in a capacity test. Policies can route simpler prompts to cheaper models, burst overflow to providers such as Bedrock and Databricks, or keep PII-tagged requests inside the cluster, and per-tenant metrics cover token rates, TTFT, ITL and cached versus uncached prompt tokens.
Why it matters: Buoyant's Kinference adds KV-cache-aware load balancing and YAML inference policies on top of vLLM, with benchmarks claiming large TTFT and goodput gains over standard Kubernetes service routing.
Clerk AI is running enterprise voice and messaging agents on LiveKit's open-source framework and global edge network, handling over 1,000 concurrent voice sessions without pre-provisioning. Agents respond in under 500ms and support 30+ languages natively, with ScreenSense call screening at 97% accuracy and TrueReach voicemail detection above 98%. The platform keeps full control over STT, LLM, and TTS providers and holds SOC 2 Type II certification with built-in TCPA compliance.
Meta introduced F3 Lightning, an infrastructure layer on top of its Facebook Feature Framework (F3) that decouples feature releases from shared execution, so each of the 20,000+ ads ranking features is compiled, validated, deployed, and rolled back independently. Comparable extraction features now reach production ~3.6x faster than prior F3 with no increase in infrastructure compute cost.
Preferred Networks will exhibit its MN-Core series at SC26, the international high-performance computing conference held November 15–20 in Chicago, with booth demonstrations running November 17–19. At booth 4606, the company will show a structural model and mockup of its in-development MN-Core L1000 inference AI chip, its future plans for the MN-Core L series, a live demo on the MN-Core 2 Devkit, and R&D on ultra-high-efficiency AI computing infrastructure.
Figma's code layers feature is now in open beta for 100% of users, letting teams build and compare code-backed prototypes directly on the Figma canvas. The walkthrough shows a designer cloning a GitHub repository into a code layer, prompting the Figma agent to generate three delivery-address variations side by side, and refining one in the full-screen code editor. Extract designs turns screens or recorded flows into editable design layers that can be sent back to the prototype with Update code layer, and the agent can merge two teammates' code layers into one combined experience.
OpenAI describes training activation classifiers, or probes, that read a model's internal residual stream activations to flag misaligned or policy-violating assistant completions, and says it has deployed them on internal and external traffic of its most capable models starting from GPT-5.6. The probes use an Attention probe architecture with MLP plus QV-attention and an exponential moving average, are trained on up to 1M messages labeled by reasoning-model judges, and run in streaming fashion on the inference engine so a completion can be blocked mid-turn. OpenAI reports excellent probe transfer across many post-training stages but more limited transfer across pre-training checkpoints, and says probes were overall more robust to jailbreaks than content-based classifiers in its red-teaming.
Why it matters: OpenAI details how activation probes on internal model activations serve as a streaming safety layer, including their limits on transfer across checkpoints and retraining needs.
Sierra has redesigned Agent Studio around Ghostwriter, its always-on agent for building agents, so investigating a problem, editing a journey, and testing a fix now happen in one conversation. Ghostwriter follows you across journeys, simulations, and dashboards, and each session keeps changes and test results together until a teammate reviews and merges them. Agent Studio also groups related agents into a single project and adds line-by-line diffs, Preview Chat, Simulations, and release governance with required approvals and controlled rollouts.
Google ran a 50-question auto-graded coding-agent benchmark testing whether a lightweight classifier, TypeSafe's Jev, can route each task to the cheapest Gemini configuration that still answers correctly. A frozen regex and prompt-length heuristic scored 93.8% on the 16 development questions but dropped 23.2 points to 70.6% on 34 held-out tasks, while Jev reached 96.0% across the full suite. The strongest no-router setup, Gemini 3.8 Flash on low thinking, solved all 50 tasks, and a Jev router sending tasks to Flash-Lite first and escalating flagged ones also solved all 50 while reducing cost and response time; capping maxOutputTokens at 300 saved nothing on easy tasks and broke 7 of 25 hard ones.
Zyphra introduced Zyphra Hierarchical Attention (ZHA), a sparse attention method that builds a tree of compressed KV-cache summaries and routes queries top-down through it, giving subquadratic complexity with suitable top-p thresholds. Existing pretrained models can be adapted via a short KL-distillation phase using only a few billion tokens; on Qwen3.6-35B-A3B, ZHA adds 81,920 trainable parameters, about 0.00024% of the base model. On RULER-13 at 256K tokens the wide-tree model scored 96.2% versus 96.0% for dense attention, and the deep six-level tree scored 63.4% on LongBench v2 versus 58.0% for dense while attending to about 1.1% of the KV cache; a GPU roofline model projects 32-35x speedup over dense attention at 10M tokens.
Ai2's Olmo 3 received Nextcloud's highest Ethical AI rating, green, which requires open-source training and inference software, a self-hostable trained model, and publicly available training data. Nextcloud bundles Olmo 3 with its llm2 local LLM backend, and the 7B variant was chosen so users without 16 or 32 GB of VRAM still get a good out-of-the-box local experience.
Anthropic released FlashProve, an open-source prototype that uses zkSNARKs to prove a response came from a committed model without revealing its weights. On Llama 4 Maverick (400B-parameter MoE, 4-bit weights), a single H100 generates proofs at about 110 tokens/second with ~19 MB proofs per 1,024-token request, roughly 70-95x the GPU time vLLM needs for the same work. The release includes executable prover and verifier code, a C++ fast verifier, a Lean verifier with machine-checked completeness, soundness and zero-knowledge statements, and a technical report, all under Apache 2.0.
Why it matters: Anthropic's FlashProve shows zkSNARK-based proofs of inference are becoming practical for a 400B-parameter MoE model, with open code and Lean-checked security claims.
Google DeepMind is expanding PlanAssist, its Gemini-built AI planning tool, to 12 additional local planning authorities in England, bringing the total to 15. In pilots across Barnet, Camden, and Dorset since June, the tool cut the average time officers spent on householder assessment tasks by up to 58%, and more than 70% of its generated reports were approved with no edits.
Toyota and Waymo announced the next phase of their autonomous driving collaboration, with initial operations to test and validate an autonomous Toyota bZ4X fleet in Japan in 2027, followed by planned deployment to the United Kingdom and other Waymo-serviced markets. Toyota will build the vehicles at its Takaoka plant in Toyota City, Aichi Prefecture, assembling the base vehicle and integrating Waymo's AV hardware under one roof. The companies describe the partnership as a milestone for Japan's first fully autonomous ride-hailing service using high-volume, domestically manufactured vehicles.
Combining knowledge distillation with the Medusa framework lets a 1.1B-parameter TinyLlama student match the quality of its 70B-parameter Llama-3.1 teacher while running 10 to 20 times as fast at less than one-third the cost. Working with Booking.com on AI Trip Planner, Smart Filters and accommodation description generation, the pipeline closed over 90% of the quality gap and added 2 to 3.6x further speedup from Medusa heads.
Meta has moved Catching JiTTests, its LLM-based just-in-time test generation approach, from research prototype into production infrastructure, and published the results in the FSE 2026 paper "Just-in-Time Catching Test Generation at Meta." The system runs overnight on spare capacity, targets risky diffs, generates tests through dodgy-diff and intent-aware workflows, and filters false positives with an LLM-as-Judge ensemble plus a rule-based RubFake assessor before reaching engineers. As of writing, 589 engineers (64% of outreach) engaged with the results, which showed a 60% true-positive rate, and in the first half of 2026 the system caught 74 serious privacy bugs at the diff stage and lowered privacy SEVs by 20%.
DevFest, Google's community-led developer event series hosted by Google Developer Groups, returns from October 1 through December 31, 2026, with 800+ events across 115 countries expected to draw almost 1 million attendees. Sessions and workshops cover Google AI technologies plus Google Cloud, Android and Web tracks, framed around building, securing and scaling in the agentic era. Events are hosted in local languages by grassroots organizers, with speakers including Google Developer Experts and Google engineers.
inclusionAI released Ling-3.0-flash-Sante-fp8, an FP8-quantized version of its healthcare-specialized Ling-3.0-flash-Sante model, keeping 124B total and 5.1B active parameters per token. It matches the BF16 checkpoint's 73.9 average across healthcare benchmarks while cutting weight memory for FP8-capable hardware. The research-only model supports 256K-token contexts and defaults to thinking mode with tool calling.
Meta detailed the next generation of Andromeda, its personalized ads retrieval engine, which scores over one billion ads per second in production — a 100x throughput lift over the 2024 system. The new architecture brings LLM-inspired mixture-of-experts to retrieval, approaching ranking-level model performance at roughly 300x fewer inference-time FLOPs, with over 10 billion ads per second on a single NVIDIA GH200 GPU. Since 2025, Andromeda has delivered a 2.4% increase in conversions on Facebook.
Replit published the architecture behind Agent Habitat, the offering that gives agents access to company knowledge and systems inside boundaries the company controls. The post sets out four principles: least authority for agents, containment of failures, accountable production changes, and governable delegated work. Replit says its existing microVM isolation prevented exploitation of recently disclosed Linux kernel vulnerabilities, and that connector credentials stay outside the agent's context, with a credential broker authorizing each request before retrieving secrets from a vault.
Zenrin, Zenrin DataCom and Preferred Networks have signed a business and capital alliance, with the Zenrin companies subscribing to a PFN third-party allotment. The partners will combine continuously updated geospatial data, including historical change records, with PFN's generative AI, physical AI and 3D spatial recognition technology to build AI models that understand and predict real-world change. Initial focus areas include public sector, real estate, retail and distribution, targeting demand forecasting, store site selection and urban planning, disaster prevention and logistics simulation.
Figma detailed how it trained a first-party design editing model for its Figma agent, starting with supervised fine-tuning on synthetic edit requests represented in Figma's native file format and validated by deterministic checks and visual model judges.
Meta detailed PrivacyLib, a common library embedded in the clients of hundreds of its data and processing systems that enforces privacy requirements at the point of data access, performing billions of checks per second. Systems integrate once, and new privacy capabilities ship as plug-in modules, replacing per-system implementations that could take months. An Agent Observability module tracks internal AI agent data access by carrying agent context to PrivacyLib.
Replit launched Cloud Agents for existing repositories, letting anyone on a team describe a change in plain language and have the agent open a tested pull request in GitHub or GitLab, available to all Pro and Enterprise customers. The agent runs each change in isolated cloud sandboxes, clicks through the running app, and attaches screenshots and video to the PR, while shared environments give every contributor a preconfigured setup that updates from the main branch. Replit also introduced early-access self-improving workspace memory, which turns review comments into team skills that engineers approve, and a Slack integration that lets users mention @Replit in a thread to start a change.
Sakana AI has opened an early access program for its Sakana API Platform, giving enterprise customers API access to its LLMs with organization- and workspace-level management of members and usage costs. The OpenAI-compatible API requires only an endpoint change, and Sakana Fugu and Fugu Ultra are available first, with Sakana Namazu and Fugu Max to follow. Features include role-based access control, CSV audit logs, budget limits, invoice payment, MSA and DPA, and a commitment not to train on customer inputs.
Google Cloud developers built HappyDelivery, a working prototype of conversational commerce using LangGraph, Gemini, FastAPI and Google Cloud, following an earlier three-layer blueprint of Ingress, Memory and Cognition. The Ingress layer validates platform-specific webhook signatures for WhatsApp, RCS, Apple Messages and SMS, normalizes payloads into a common schema and offloads them to Pub/Sub, returning HTTP 200 in roughly 300ms. The post walks through the code, state machine, deterministic financial tools and an automated verification suite.
Google released swappable inference formats for A2UI, its agent-optimized protocol for generative UI, letting agents generate an intermediate format that SDK compilers turn into validated A2UI messages. The new default, A2UI Express, uses up to 3x fewer tokens than Direct JSON for the same UI and cuts latency up to 3x, and it ships in the Python (a2ui-agent-sdk 0.9.0), TypeScript (@a2ui/agent) and Dart SDKs. Google also showed decision models such as TypeSafe's Jev choosing among predefined layouts, which it says can reduce latency up to 80x; in the Apparite example, screens were complete in 0.8 to 1.7 s across 11 prompts.
Preferred Networks began offering PLaMo 3.1 Prime on October 27, 2026, adding Web search and a Deep Research feature that autonomously plans, collects, analyzes and writes source-cited reports over several minutes to tens of minutes. Deep Research also runs on-premises and can connect to internal document search systems, while the company reports an average ~15% score improvement across eight hallucination-related benchmarks including AA Omniscience versus PLaMo 3.0 Prime. PFN also cites 1.7x faster processing via speculative decoding and a 90% reduction in output tokens in non-reasoning mode, and launched a free PLaMo API plan plus a PLaMo Chat campaign offering Deep Research twice daily until November 30, 2026.
Why it matters: PFN's PLaMo 3.1 Prime adds autonomous Deep Research that runs on-premises, giving Japanese enterprises a locally deployable option for source-cited investigation work.
Apple introduced updates to Apple Creator Studio that bring an intelligent assistant built on Apple and third-party models to Keynote, Pages, and Numbers, plus Model Context Protocol support so external AI agents can read content and perform edits in those apps. Final Cut Pro for Mac now includes built-in MotionVFX and Content Hub stock video, Logic Pro gains a beta assistant and a Stem Splitter that separates drums, and Pixelmator Pro adds Generate Fill for natural-language image edits. The updates arrive on the App Store this November, free for existing subscribers, with new subscriptions at $12.99 per month or $129 per year.
Replit predicts every Fortune 500 company will become a "Self-Driving Company" by the end of 2030, with agents executing and coordinating work while people set goals and remain accountable. Replit says its engineers nearly tripled code output in the first half of 2026 while review times, reversions and product incidents stayed flat. Replit for Business is being built as the foundation, and at SMFL a development process that previously took six people six months was completed by one person in one week.
Vapi has launched traffic splitting, letting teams route 0.1%, 1% or 5% of inbound and outbound calls to a new agent version before committing. Previously, publishing a new assistant version sent 100% of traffic to it immediately. The feature includes "stickiness" to keep repeat callers on the same version, and Vapi recommends starting with two-way splits.
Mastra is opening its Knowledge service, a self-hosted company-wide context layer for AI agents, to eight more design partners on a first-come, first-served basis. Knowledge connects systems like Slack and Monday, keeps itself current, and supports separate knowledge bases per function with permissions at the individual, team and org level. Mastra says wiring a company-wide knowledge graph into its coding agents produced a noticeable improvement in shipped code quality.
Google Research described at IDWeek 2026 how its chest X-ray TB screening models progressed from a 2019 Radiology benchmark to a CE MDR Class IIb certified tool deployed by partner Nexus AI in about 40 healthcare facilities across six countries. More than 90 clinicians have evaluated over 25,000 chest radiographs, cutting turnaround from weeks to under 45 seconds, and the tool has also been deployed offline four kilometers underground at South Africa's Mponeng Gold Mine. Google has open-sourced chest X-ray embeddings and MedSigLIP through HAI-DEF and released MedGemma, which it says has passed 10 million downloads, while Indonesia's Ministry of Health is fine-tuning both models on domestic radiographs for a national screening program.
Amazon released TorchNeuron, an open-source PyTorch backend that translates PyTorch code into operations executable on AWS Trainium systems with minimal or no code changes. It adds an opt-in adaptive eager mode that groups and compiles operations for near-compiled performance, plus multi-level NEFF caching; a shared compilation service cut time to first training step by about a third in a 64-process LLM test. The Annapurna team tuned and benchmarked it using 841 representative models distilled from hundreds of thousands of open-source models.
Modal's forward deployed engineering team worked with Sakana AI to run Sakana Namazu, a ~1T-parameter MoE model tuned for Japanese business context, on Modal for its public API launch. The team applied selective NVFP4 quantization to routed experts only, keeping attention, the router and shared experts in BF16, and calibrated on Japanese-heavy conversations so all 23,040 experts saw representative traffic. A custom DFlash draft model trained on more than 30 billion raw token IDs raised speculative token acceptance by 25% and cut per-token decode latency by about 35%, while a second drafter trained on roughly 1M real requests lifted average acceptance length from 3.51 to 5.01 and cut p99 latency from 68s to 52s.
Figma agent is now generally available, and Figma explains how to feed it your design system so its output matches your product. Connecting a published library lets the agent use approved components and tokens, while guidelines files encode team rules and Examples pages or _example naming supply reusable patterns.
Sakana AI announced Sakana Aquarium, a platform for building, running and operating AI agents that support organizational decision-making and execution on top of an ontology. The platform combines three parts: an Ontology that organizes knowledge, rules and data with their meanings and relationships while connecting existing sources without migration; Governance that sets access permissions and pre-execution approvals and logs everything from decision rationale to execution results; and Managed Agents that deploy workflows covering information lookup, AI judgment, function execution and human approval. Sakana AI says the data connections and ontology can be reused across agents, and that the platform grew out of its FDE-led work with financial institutions such as MUFG, SMBC Group and Daiwa Securities Group.
Modal released Spindle, an open-source, Tinker-compatible training server that runs in a customer's Modal workspace, letting teams customize the stack and pay for GPU time instead of tokens. Spindle supports full-parameter training and multi-LoRA across open models, and Modal reports that sharing one trainer across 8 clients cut cost per 1M output tokens by up to 80% versus Tinker on a DAPO-math GRPO job. Modal says Spindle is still in beta and positions it between its Dojo training library and raw cluster primitives.
Why it matters: Modal's open-source Spindle server offers teams a Tinker-compatible alternative that trades per-token pricing for GPU time, with multi-LoRA sharing reported to cut costs by up to 80%.
Google Cloud researchers unveiled Project Chrysalis, a reference architecture for governed, self-evolving AI agents built on the Gemini platform and Agent Development Kit (ADK). In simulated Black Friday cache-stampede outages, it cut time-to-first-mitigation from 15–45 minutes to 1.8 seconds, reduced HTTP 5xx errors from 72.7%–80.0% to 0.0%, and dropped P99 latency from 519.8 ms to 18.2 ms. Every autonomous action is bounded by Model Armor, Cloud KMS, Binary Authorization, OPA Policy Controller and dual-custody human-in-the-loop sign-off.