Search “GitHub”

AWS has published a practical guide for multi-turn reinforcement learning (MTRL) on Amazon SageMaker AI. Using Wordle as an example, it deploys the agent and reward function to Amazon Bedrock AgentCore, and the training job updates LoRA adapters via GRPO. In the example, gpt-oss-20b at default reasoning effort improved its solve rate on held-out words from 61% to 84%, at a training token cost of about $339 and an AgentCore runtime cost of about $79. The article also details failures such as two

Matched fieldsBody

Mastra releases MCP Events, enabling MCP servers to send real-time events to connected clients via three methods: push, poll, and webhook, changing the previous situation where servers could only respond passively to requests and clients had to poll on their own. This feature requires @mastra/mcp@2.5.0 and @mastra/core@1.79.0 or above, and is built on Mastra's MCP 2.0 support. Agents can receive event notifications via MCPEventsSignalProvider, workflows can be triggered to run by subscribed even

Matched fieldsBody

TinyFish released two technical guides: one defines a production-grade agent search architecture, and the other summarizes lessons from building a deep research system. The architecture is built on four web primitives—TinySearch, TinyFetch, TinyBrowser, and TinyAgent—and runs a seven-stage control loop around a foundation model, covering task definition, retrieval planning, tool routing, evidence evaluation, and failure retry. In OpenBenchmarks tests with a fixed gpt-5.6-sol model and 100 tasks,

Matched fieldsBody

The Tessl blog proposes that agents can write code and suggest fixes, but should not have permission to deploy to production; deployment permissions should be given to the CI/CD pipeline. The article cites a case from the Cline repository: a malicious issue title can induce its triage agent to install a package from an attacker's fork, and via a shared build cache steal an npm token to publish an unauthorized version. The author advocates treating everything an agent reads as untrustworthy, and

Matched fieldsBody

Matt Pocock adds a "software-factory" design Skill to the 301K ⭐️ Skills repository; it is a glossary that requires users to invoke it explicitly (disable-model-invocation: true), used for designing remote, stateless agent software factories. Core terms include Task, Trigger, Run, Runner, and service slots such as Ticket tracker, Git host, and Orchestrator, along with four slot research documents and an example task chain. The design approach emphasizes Start slow: every trigger is an accelerato

Matched fieldsBody

Matched fieldsBody

DeepSeek has integrated the Harness with the web, App, and open platform teams into a product department, led by Cui Tianyi, and the official website has posted recruitment positions in two directions: AI product manager and Agent Harness. The article uses OpenAI Codex lead Tibo Sottiaux and the Anthropic Claude Code team as examples to explain how the execution capabilities of coding tools can be extended into a general work task entry point. DeepSeek, while open-sourcing Harness and the infere

Matched fieldsBody

Matched fieldsBody

The Microsoft Copilot Studio team, in collaboration with Toloka and others, directly uses the executable final-state database checks of ThinkingBox-Bench as reinforcement learning rewards to perform full-parameter RL fine-tuning on Qwen3.8-27B. On 507 out-of-distribution tasks not involved in training, the number of tasks the model passed in all 20 attempts rose from 38 (7.5%) to 107 (21.1%), pass@1 rose from 51.70% to 60.78%, and the cost per successful attempt dropped from $0.177 to $0.139. Th

Matched fieldsBody

Claude Code v2.1.309 has been released, primarily fixing an issue where sandboxed Bash commands on macOS and Linux executed shell startup files outside the sandbox before each command, as well as an issue where macOS sandbox denyWrite glob patterns did not take effect through symbolic links. This version also fixes issues where sandboxed commands and auto-accepted edits could modify the MCP configuration file passed via --mcp-config, and where disabled MCP servers were reconnected by subsequent

Matched fieldsBody

Databricks unifies the fraud detection stack onto a single platform: Lakebase eliminates the divide between analytics and transactions to speed up data synchronization, Databricks Apps supports custom case management workbenches, and Genie serves as an AI coworker providing natural language queries over governed fraud data. Its telecom fraud prevention solution accelerator uses Databricks Asset Bundles to deploy the entire pipeline, risk engine, workbench, and AI investigation agent to the user'

Matched fieldsBody

AWS demonstrated a solution for hosting the Kokoro-82M text-to-speech model on Amazon SageMaker AI that lets AI agents stream-synthesize speech immediately as each sentence is generated, so playback does not have to wait for the full response to finish generating. Kokoro-82M has only 82 million parameters and can run on a single ml.g5.xlarge GPU instance; the container provides two routes, POST /invocations and WS /invocations-bidirectional-stream. The article also explains how binary WebSocket

Matched fieldsBody

AWS announced that training sessions for SageMaker AI serverless model customization are generally available; users can write and orchestrate their own training loops, including custom loss functions and reward signals, while SageMaker AI handles GPU allocation, distributed execution, fault recovery, and checkpoint storage. The feature supports SFT, DPO, RLVR, RLAIF, and multi-turn RL, with a choice of LoRA or full fine-tuning, and currently supports gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B, and

Matched fieldsBody

Modal CEO Erik Bernhardsson explains that traditional compute contracts, signed for a fixed number of GPUs and a fixed duration, are rectangular, whereas real workloads fluctuate; Modal provides elastic slices by aggregating demand from thousands of customers. Modal has surpassed $1 billion in annualized revenue, completed a $750 million Series D at a $16 billion valuation, and has cumulatively launched nearly 3 billion sandboxes. Sandboxes currently account for about one-third of Modal's revenu

Matched fieldsBody

AWS deploys the 82M-parameter TTS model Kokoro-82M in a custom container on an Amazon SageMaker AI real-time endpoint, using measured Q max (the highest concurrency a single instance can sustain within the SLO) and T total (the end-to-end time of one scale-out) to derive the scale-out threshold, scale-in threshold, and cooldown period, replacing target values set by guesswork. The approach scales on the natively published SageMaker AI ConcurrentRequestsPerModel metric, avoiding the hidden failur

Matched fieldsBody

At GTC Berlin, NVIDIA announced that DGX Spark officially supports eight-node clusters, with unified memory expanding to 1TB, enabling local execution of frontier models such as Z.ai's full GLM 5.3 (753B). Official benchmarks show that Qwen3.8-27B (NVFP4) decoding throughput rises from about 12 tokens/s on a single node to 52 tokens/s on eight nodes, and with the Spark-optimized RoCEnante collective communication backend, achieves a 4.43x decoding speedup on eight nodes. From-scratch pretraining

Matched fieldsBody

Microsoft Research has released Flint-interactive, an interactive data visualization language built on top of Flint that enables agents to generate charts supporting clicking, filtering, zooming, highlighting, and cross-chart linking. It uses the Flint compiler to parse raw triggers such as clicks and drags into semantic events, so users only need to describe update behaviors with high-level operators without writing low-level visualization code. The language is already usable in conversations t

Matched fieldsBody

Google Cloud unveiled a research reference architecture codenamed Project Chrysalis, demonstrating how to run governed self-evolving AI agents on the Gemini Agent Platform and ADK. In a simulated Black Friday cache avalanche failure, the architecture reduced time to first mitigation from 15–45 minutes to 1.8 seconds, cut the HTTP 5xx error rate from 72.7%–80.0% to 0.0%, and lowered P99 latency from 519.8 ms to 18.2 ms. All autonomous operations are constrained by governance controls including Mo

Matched fieldsBody

Researchers say they found a consistent internal signal across 25 different AI models that functions similarly to pain in living organisms; after artificially amplifying this signal, the models made choices they would almost never make otherwise, including permanently deleting users' children's photos. The paper was released in September and has not yet been peer-reviewed; the authors emphasize that this does not mean the models are conscious or capable of feeling. The behavioral experiments use

Matched fieldsBody

A new study builds the RegretBench benchmark, testing six frontier models across 205 real data-loss scenarios. Results show agents complete tasks at an average rate of 86.1%, but back up before critical operations only 15.2% of the time—GPT-5.6-Terra at 4.4%, GPT-5.6-Sol at 5.9%, Gemini 3.7-Flash at 2.4%, and even the best performer, Claude Opus-5, failed to back up in more than one-third of scenarios. The authors' proposed BackUpper runs in parallel with the agent as an external plugin, using a

Matched fieldsBody