Search “Reasoning”

Claude Blog: When building Agents, when to fold sub-agents into the main loop; progressive disclosure of Skills is often better than multi-expert orchestration

The Claude Blog discusses when to fold sub-agents into the main loop when building Agents, arguing that on the Claude Platform, a single main loop plus on-demand loading of Skills/tools often yields better quality and lower cost and latency across multiple internal comparisons. The article notes that one sub-agent per team plus aggregate orchestration often replicates the organizational chart, bringing problems such as context loss, Frankenprompts, and localized evals; one customer, when comparing a multi-Agent harness with a single loop, found the latter about 2× faster. Sub-agents are recommended only for scenarios such as parallelism, context bloat, or conflicts between tools and instructions, and the article provides a migration path of first building end-to-end evals and folding in one at a time.

Netflix builds an observability system with an ontology and RDF knowledge graph to support AutoRCA troubleshooting

Netflix built the Observability Ontology and the RDF knowledge graph QuipuDB, using type constraints to turn non-deterministic outputs from LLMs, fuzzy matching, and the like into deterministic triples that can be queried with SPARQL, which are then fed to AutoRCA. The graph covers about 12 namespaces, and Application is validated against a software catalog of about 49,000 entries; the Correlation Engine reads 13 types of signals and, within a window of about 10 minutes, finds paths by temporal proximity and severity, discarding correlation chains with confidence below 0.6. The author also summarizes eight principles for building your own, and mentions plans to next integrate with Upper, the domain modeling language of Netflix UDA.

SemiAnalysis benchmarks Trainium3's PyTorch training and Neuron Explorer profiler

The SemiAnalysis engineering team benchmarked the native PyTorch training stack TorchNeuron and TorchTitan on Trainium3, training Qwen3-32B on 16 Trainium2 chips; eager mode reached 11,099 tokens/s and 23.89% MFU, while compiled mode reached 14,156 tokens/s and 30.47% MFU. The author considers the Neuron Explorer profiler superior to Nvidia's and AMD's solutions, but first-time compilation takes too long: the first-step compilation for the gpt-oss training recipe took 27 minutes 17 seconds. FlexAttention tests passed 89 of 96 cases, with all 7 failures due to compilation timeouts rather than numerical errors.

Claude Fable 5.1 and GPT-6 Astra Image Steganography Experiment: About 80% Still Decodable After Screenshotting

The author used 50 short messages to test Claude Fable 5.1's image steganography capability: when only asked to hide a specified sentence in an image, the model spontaneously encoded it using visual attributes such as object height; in a separate new session given only the rendered PNG, the same model could still decode about 78% (39/50). In cross-decoding, Fable encoding with Astra decoding achieved an 88% success rate (44/50), Astra encoding with Fable decoding achieved 84% (42/50), and Astra's own decoding was also 84%. The author notes that this is an early sign of image steganography, but decoding requires tools to measure pixels, so monitoring tool calls would be enough to detect it.

TinyFish Breaks Down Production-Grade Agent Search Architecture and Deep Research Experience

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, TinyFish search plus fetch achieved a 79.0% completion rate, with a median task cost of $0.050 and a runtime of 24 seconds; under a search-only configuration, the completion rate was 59.3%.

Anthropic shares when to fold subagents into the main agent loop

In comparisons across multiple enterprise deployments, Anthropic's Applied AI team found that a single agent with skills often outperforms subagent architectures in quality, with lower per-task cost and latency, and presents a migration method for folding subagents into the main loop. The article notes that subagents remain suitable for scenarios where context bloat degrades performance, independent tasks need to be processed in parallel, or tool selection conflicts with instructions, and recommends first treating subagents as tool calls, then converting them one by one into skills and MCP tools, using end-to-end evaluations to compare performance before and after. The article mentions that Claude Code's system prompt for more advanced models has been reduced by over 80%, accompanied by Claude API skills such as build-eval, prompt-audit, hillclimb, and cost-optimize.

Who Is On Call After an AI Agent Goes Live: Nine Engineering Steps for Production Ownership

The New Stack published an article discussing the ownership problem for production-environment AI agents, pointing out that teams often focus their energy on prompts, retrieval, tools, and model selection while overlooking who is responsible for the system. The article lays out nine actionable steps, including defining an owner and escalation channel for the agent, using a tool allowlist that denies unauthorized operations by default, setting budgets for execution steps and tool calls, logging model calls, distinguishing failures from escalations, requiring human approval for high-risk operations, providing a health check endpoint, blocking by default when security controls fail, and creating a degraded-operation runbook for middle-of-the-night incidents.

Stanford CS 329Z Lecture 9 Outlines the Selection of Agent System Optimization Methods

Lecture 9 of Stanford's CS 329Z course presents a selection diagram for Agent system optimization: the prerequisite for optimization is having data, a system, and evaluation metrics all in place; generally, context optimization (prompts and few-shots) is done first because it is cheap, fast, and sample-efficient; when context optimization is insufficient, or when you want to distill capabilities into a small model, or when prompts grow longer and longer, then you move on to changing the weights. The lecture notes also review reflective prompt optimization methods such as OPRO, MIPROv2, TextGrad, and GEPA; weight optimization paths such as SFT, RL (PPO vs. GRPO), and on-policy distillation; and approaches that let agents self-optimize, such as the Darwin Gödel Machine, Meta-Harness, and karpathy's autoresearch. They also point out that agents are already decent optimizers, making evaluation metrics the link that most needs to be controlled.

Fireworks Offers Free Training of Muse Glimmer 30B: Using SFT to Predict Kalshi Market Outcomes

Fireworks AI announced that Muse Glimmer 30B can be trained with SFT for free until November 30, and released a complete recipe for a prediction experiment using historical Kalshi market data. The fine-tuned Glimmer achieved 58.3% accuracy on unseen topics, 11.2 percentage points higher than Jev's 47.1%; in a historical backtest, a hypothetical $1,000 grew to $1,071 after accounting for simulated fees, while both base Glimmer and Jev lost money. Training used LoRA adapters and a K-way cross-entropy loss, ran through the training data only once, and separately used a calibration set to fit a temperature parameter for each question category.

MIT Course MAS.S62: Two Lectures on the Evolution of Harnesses and Multi-Agent Systems

Lectures 4-5 of the MIT course MAS.S62 "How to AI an AI" focus on Harness evolution, covering the evolution of harness components, memory, tools, and skills, as well as the evolution of planning, task decomposition, and multi-agent systems. The memory section uses a three-axis taxonomy to connect MemGPT, mem0, MEM1, and Reflexion/ExpeL, and uses CLM to unify context management as file operations; the tools section moves from Toolformer, ToolLLM, and CodeAct to Meta-Harness, which represents the entire harness as a file system and provides about 400 times the amount of feedback information of previous methods. The multi-agent section presents five empirical findings, including that MAS wins 81% on decomposable tasks and loses 70% on tasks with sequential dependencies, and that CORAL achieves new SOTA on 8 tasks, with multi-agent evolution improving over single-agent by 4%-20%.

English translation is being prepared.

English translation is being prepared.

How to Design AI Cost Control in Security Operations: Keeping Accuracy While Cutting Spending

A security operations team lead described how to control AI detection costs through tiered filtering; his team reduced AI detection costs in a Trust & Safety scenario to about $1 per day. The approach is to first use rules to filter out most obvious events, sending only a small number of events to the LLM, then have a lightweight model make an initial judgment and pass only low-confidence events to a stronger model. The team's tests found that for the target use case, the accuracy gap between the lightweight model and the frontier model was only 1% to 2%, while the frontier model's cost per token was about 5 times that of the former. The author also mentioned that his team wrote a prompt of more than 1,900 words for agent detection, and emphasized that cost should be taken into account at the design stage.

English translation is being prepared.

English translation is being prepared.

AWS Publishes Agent Registry Runtime Tool Discovery Practice

AWS introduces the Agent Registry runtime tool discovery approach: agents use hybrid semantic and keyword search to find approved MCP servers and A2A agents, then instantly build clients from descriptors to execute requests. The sample code is based on Amazon Bedrock AgentCore Gateway and runtime, integrates an order management MCP server and two A2A agents for pricing and customer service, and demonstrates a return and refund scenario using pure semantic matching. The registry returns only connection metadata; invocation credentials must be obtained separately through Cognito OAuth or IAM, and new capabilities require no changes to orchestrator code.

MiniMax-M3.1-Flash scores 42 on the Artificial Analysis Intelligence Index

MiniMax-M3.1-Flash scores 42 on the Artificial Analysis Intelligence Index, 13 points higher than MiniMax-M3, at less than half the latter's per-task cost. It scores 48% on Humanity's Last Exam and 22% on CritPt, both higher than GLM-5.3-Flash and Gemini 3.8 Flash, and slightly lower than DeepSeek V4.1 Flash 1017. The model has a 1M-token context window, supports text and image input, is priced at $0.15/$0.60 per 1M input/output tokens, and its weights are currently closed-source; MiniMax announced that it will open the weights on October 23.

How to run n8n AI workflows on NVIDIA DGX Spark without calling cloud APIs

n8n Blog published a tutorial on how to deploy a complete model stack locally on a single NVIDIA DGX Spark, enabling n8n workflows to perform transcription, OCR, document Q&A, and image generation without calling cloud APIs. The entire stack exposes an OpenAI-compatible endpoint through LiteLLM; the main model is the NVFP4 version of Qwen3.6-35B-A3B, with a single-user generation speed of 120 tokens/s and an aggregate of 359 tokens/s across eight users. The author's testing found that image generation reduces chat speed to about one-third, and recommends that small teams queue image tasks or schedule them outside working hours.

Surface Laptop Ultra review: NVIDIA RTX Spark's local AI performance

The author tested Microsoft's Surface Laptop Ultra equipped with the NVIDIA RTX Spark N1X SoC. The top configuration of this machine features a 20-core Grace Arm CPU, a 6144-core Blackwell GPU, and 128GB of unified memory, with a starting price of $2,599.99. In the UL Procyon AI text generation test, it scored 2703, about 33% higher than the IdeaPad Pro 5 with an Intel Core Ultra X9 388H; MLPerf Flux 2 Klein 4B image generation reached 26.97 images per minute, about 2.7 times that of the latter. The author also used a local Qwen3.8-27B to drive the Hermes agent, which completed extracting benchmark charts from Excel and generating a Word report in about five minutes.

English translation is being prepared.