LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-11-1298 publications

Actualiser

Alibaba Cloud Agentic SOC Team: Organizational Restructuring for Multi-Agent Collaboration Reduces Delivery Cycle from 51 to 14 Days

Matériel
Vérifié 2026-11-12 01:58 GMT+8

As AI agents take on more execution steps, the next challenge facing teams is not tool capability but the collaboration structure itself. When developing this security operations product, the Alibaba Cloud Agentic SOC team attempted to embed multi-agent collaboration into organizational processes and documented a specific restructuring methodology.

The starting point of the problem is the inherent delay of sequential handoffs. In traditional models, product teams complete requirements, security teams finalize rules, engineering teams implement solutions, and testing teams conduct acceptance; each stage produces a deliverable, yet the customer's issue is not considered resolved until the entire chain has run its course. Accelerating any single link yields limited total time reduction—goals are repeatedly reinterpreted during handoffs, different functions queue separately, acceptance is concentrated at the end, and problems travel back along the same path.

The team illustrated the difference using a schedule comparison for one requirement. Under the sequential handoff model, requirement clarification, design, security standards, backend development, frontend development, and centralized acceptance proceeded sequentially, totaling 51 days. The alternative approach established 12 acceptance criteria before task commencement, appointed one person fully accountable for the result throughout the process, and organized product experience, engineering delivery, and security effectiveness as parallel tasks, shortening the cycle to 14 days.

What merits attention in this comparison is not the number of days themselves, but two structural changes: moving acceptance criteria from the end to the start, and shifting accountability from stage-by-stage transfer to permanent assignment to one individual. The former ensures all functions work toward the same standard from the outset; the latter eliminates the loss caused by goals being repeatedly reinterpreted during handoffs.

To enable agents to advance autonomously within a parallel structure, the team defined four types of boundaries. The goal boundary answers what problem may be solved—the task objective, priority, input scope, and expected deliverables must be explicit, requiring renewed confirmation for work beyond that scope. The quality boundary answers what state constitutes completion—the original issue must pass retesting, no regression may appear within the agreed evaluation scope, and no definitive conclusion may be produced when evidence is insufficient. The permission boundary answers what may be accessed and changed—available data, tools, and environments must be restricted, high-risk operations must remain within the scope of prior authorization, and production changes must be traceable and reversible. The stop boundary answers when a task must be handed over to a person—an anomaly, unauthorized action, deterioration in a key metric, reaching a retry or resource-consumption limit, or an inability to continue making reliable judgments should all trigger a pause.

Sources :Alibaba Cloud Community

Recherche

ElevenLabs Reveals Enterprise Voice AI Deployment Details: Dutch Fine-Tuning Boosts Identity Verification Accuracy to Approximately 90%

Prise rapide
2026-11-11 20:00 GMT+8

Voice AI is entering daily enterprise operations, where success depends not just on model capabilities but on integration around specific business scenarios. On November 11, 2026, ElevenLabs published an article detailing enterprise customer cases, disclosing deployment specifics for six companies: Yahoo, Westgate, KPN, Doppel, Pocket, and Twilio. The most operationally informative case is KPN's adaptation for the Dutch language.

KPN, the largest telecommunications provider in the Netherlands, sought to embed voice AI into services used by millions of consumers and small businesses daily. ElevenLabs noted that Dutch presents unique challenges for voice models: dense compound words, strong regional accents, and crucially, details like postal codes and dates of birth—which are exactly what transcription models struggle to process consistently. ElevenLabs' engineering team worked directly with KPN to fine-tune the model on Dutch language data, raising identity verification accuracy from roughly 50% to 90%. ElevenLabs stated that all subsequent Dutch language deployments on the platform have benefited from this improvement.

Other cases illustrate similar integration logic. Westgate Resorts, which hosts over two million owners and guests annually across nearly 60 resorts in North America, developed 'Sarah,' a front-desk voice agent using ElevenAgents. It enhanced Sarah's knowledge base with 8,000 real customer service call transcripts, ensuring responses reflected questions guests actually ask. Yahoo, meanwhile, leveraged ElevenAgents to build voice capabilities directly into Yahoo Scout, avoiding the need to construct voice infrastructure from scratch and enabling a rapid launch of real-time voice conversation features; before exiting its beta phase, Scout had delivered over 3 billion answers.

These cases collectively demonstrate that practical enterprise voice AI deployment relies on specific operational conditions: the volume of real business data used for training, the depth of engineering adaptation for target languages, and the method of integration with existing infrastructure. ElevenLabs self-reported these figures and results without providing independent verification methods or third-party audits. For readers evaluating voice AI adoption, these insights offer concrete integration references, but key metrics such as the magnitude of accuracy improvements remain vendor claims.

Sources :ElevenLabs

Recherche

Google Cloud Previews ai.classify and ai.extract SQL Functions in AlloyDB AI

Matériel
2026-11-12

Google Cloud announced on November 12, 2026, the preview of two new SQL functions, ai.classify and ai.extract, within its AlloyDB AI query engine. These functions allow developers to perform text classification and structured data extraction directly inside the database using standard SQL, eliminating the need to export data to external model endpoints.

Historically, processing vast amounts of unstructured text stored in transactional databases—such as clinical notes, customer reviews, and support tickets—required building complex application code. Developers had to export data, call external APIs, write custom prompt wrappers, parse unpredictable responses, and manage retry logic. This application-side loop introduced query latency, bloated application code, increased model token costs, and created data compliance risks by moving sensitive information across multiple network boundaries.

The new ai.classify function supports single-label and multi-label classification, allowing users to define categories via simple label lists or 2D arrays containing descriptions, and supports few-shot prompting. The ai.extract function transforms unstructured text into machine-readable JSON data matching a specified schema, handling system instructions internally to save developers from writing prompt templates or response parsers.

To enhance performance, Google Cloud introduced 'smart batching.' When these AI functions are used in SQL WHERE clause expressions, the database engine automatically groups inputs into optimized micro-batches before sending them to Gemini Enterprise Agent Platform models. This mechanism significantly speeds up inference compared to row-at-a-time evaluations and reduces total model token spend by optimizing input prompt overhead.

These features are currently in preview and require the google_ml_integration extension (version 1.6.1 or later) on PostgreSQL 18, with the enable_preview_ai_functions parameter set to TRUE. Google Cloud plans to further expand in-database AI capabilities with upcoming enhancements for advanced cost optimization, line-level provenance tracking, and query pushdowns beyond WHERE clauses.

Sources :Google Cloud Blog

Recherche

aiXcoder Open-Sources 4B Context Compression Model, Claims 78% Lower Inference Cost vs DeepSeek-V4-Pro

Sujet · 硅心科技上下文压缩模型Matériel
Vérifié 2026-11-12 01:33 GMT+8

On November 11, 2026, aiXcoder, an AI programming startup incubated from Peking University's Institute of Software Engineering, officially open-sourced aiXcompact-4B, a dedicated context compression model. With only 4 billion parameters, the model is designed for long-horizon complex development tasks involving coding agents and supports private deployment on a single A100-40GB GPU.

As AI coding tools like Claude Code and Codex handle increasingly complex tasks, agents must continuously read code, call tools, and run tests, causing token consumption to surge. Research based on the OpenHands framework found that a complete agentic coding task consumes an average of 4.17 million tokens, 3,500 times more than single-turn code reasoning. aiXcompact-4B aims to alleviate this cost pressure by structurally processing the agent's historical trajectory: it retains key original information, condenses tool return results such as file reads and test runs, removes redundant or obsolete intermediate records, and generates structured task summaries.

According to self-reported data from aiXcoder, when integrated into OpenCode for context compression during Terminal-Bench evaluations, aiXcompact-4B achieved overall task completion accuracy comparable to DeepSeek-V4-Pro. In private deployment scenarios, the single-task GPU inference cost was reported to be 78% lower than DeepSeek-V4-Pro. For API calls, the estimated token cost per request is approximately 0.019 RMB, representing reductions of 97.30%, 97.08%, and 97.98% compared to reference costs for DeepSeek-V4-Pro, GLM-5.3, and Qwen3.8-Max, respectively. Additionally, the model compresses historical context to an average of 25.89% of its original length, freeing up more context space for subsequent agent calls.

Technically, aiXcompact-4B combines teacher large model distillation, LoRA fine-tuning of a 4B student model, and structured compression decisions. Training data is derived entirely from real-world agent execution trajectories in development scenarios. While the vendor claims performance metrics close to or partially exceeding the larger DeepSeek-V4-Pro in rule-based scores and retention rates, these results are based on internal evaluation sets and specific configurations (such as ngram k=32) and have not been independently verified by third parties. For enterprises seeking to reduce high-frequency context compression costs, this offers a lightweight and economical potential choice, though actual benefits depend on specific workloads.

Sources :GitHub

Recherche

Databricks Announces GA of Genie Ontology, Citing 91% Agent Task Pass Rate in Benchmark

Matériel
2026-11-11 23:00 GMT+8

US data intelligence company Databricks announced on November 11, 2026, that its "Genie Ontology" is now Generally Available (GA). This shared knowledge layer aims to solve the "context gap" in enterprise AI agents by automatically learning from daily work artifacts and allowing human experts to govern critical business definitions.

In a benchmark developed with evaluation firm Turing, simulating a realistic UK energy supplier workspace, Genie powered by Genie Ontology achieved a 91% pass rate on enterprise task runs. Two leading frontier agents scored 45% and 58%, respectively. The Genie agent also used less than half the tokens and completed tasks three to four times faster. Databricks attributes this difference to context efficiency: Genie draws on pre-learned business rules, whereas frontier agents must rediscover knowledge during each run.

The release includes Genie Search for connecting external document sources and exposes the same context to external agents like ChatGPT, Claude, and Microsoft Copilot via Genie MCP. While the benchmark results are self-reported by the vendor, the architecture addresses fragmentation and conflict in multi-agent deployments, potentially influencing how enterprises scale AI agents across functions.

Sources :Databricks

Recherche

Google Cloud Launches Gemini for Healthcare: Cross-System EHR Access and Insurance Review Support

Sujet · 谷歌云医疗版GeminiMatériel
2026-11-11 18:12 GMT+8

US cloud provider Google Cloud announced on November 10, 2026, the launch of "Gemini for Healthcare," a specialized AI solution extending its existing Gemini Agent capabilities. Rather than merely answering questions, this tool enables AI agents to retrieve data across disparate systems and assist in completing multi-step medical and administrative workflows. The solution is currently available in preview in the United States for Gemini Enterprise Standard and Plus customers.

To address the long-standing challenge of fragmented medical data spread across patient records, insurance claims, and administrative systems, Gemini for Healthcare introduces medical-specific Skills (work guidelines) and data connectors. Defined by medical experts, these Skills instruct the AI to gather necessary information according to hospital or insurance company operational rules. For instance, before a consultation, an AI agent can automatically review a patient's past diagnoses, test results, and medication records, summarizing current health issues, allergies, and pending tests to help physicians quickly grasp the patient's status.

In terms of system integration, the solution provides data connectors for Epic and Oracle Health electronic health record (EHR) systems, as well as InterSystems' FHIR MCP Server. To preserve the integrity of original medical records, the Epic and Oracle Health connectors currently allow AI to only read EHR data, strictly prohibiting any direct modification or write-back operations. Additionally, the AI can query external resources such as disease codes, drug information, clinical guidelines, and insurance coverage rules.

Beyond clinical support, the solution targets common US healthcare insurance processes like prior authorization and claims review. The AI agent can extract relevant evidence from medical records, cross-reference it with insurance policies to organize application materials, and assist in drafting appeal documents if an application is denied. However, the AI primarily handles the collection, organization, and comparison of information; final clinical and insurance review decisions remain the responsibility of qualified professionals.

Currently, US healthcare groups HCA Healthcare and Northwell Health, along with health services provider Optum, are exploring related applications in partnership with Google Cloud. Although Michael Schlosser, Chief Medical Officer at HCA Healthcare, noted that the tool reduces the time physicians spend gathering information, Google Cloud has not yet published specific quantified results regarding time savings or efficiency gains. The actual extent to which it alleviates the burden on healthcare workers remains to be validated through real-world implementation.

Sources :iThome

Recherche

Sentence Transformers Maintainer Tom Aarsen: Pooling-Aware Training Preserves 97.8% Quality Under 4x Compression for Multi-Vector Embeddings

Matériel
2026-11-11 08:00 GMT+8

Multi-vector embedding models, such as ColBERT, provide each query token with its own independent opportunity to find a match within a document. This makes them highly expressive for search, but the cost is that each document generates hundreds of vectors. A corpus containing millions of documents may hold hundreds of millions of vectors; even with 4-bit residual quantization via Fast-PLAID, 200 vectors of 128 dimensions require 12.8 KB per document, totaling 12.8 GB across a million documents.

Token pooling compresses representations by merging similar token vectors through Ward hierarchical clustering. For example, with pool_factor=4, a document retains only about one-quarter of its original vectors. The more aggressive the compression, the easier it becomes to lose discriminative information critical for retrieval—a risk because standard training objectives never required the encoder to adapt to such compression.

In a blog post published on November 11, 2026, Sentence Transformers maintainer Tom Aarsen introduced and implemented a pooling-aware training method. Its core concept resembles Matryoshka embeddings: applying retrieval objectives simultaneously across multiple representation granularities. Specifically, each training batch contributes two losses—the retrieval loss for original document vectors and the retrieval loss for pooled document vectors—averaged with equal weights of 0.5.

In implementation, MultiVectorTokenPoolingLoss wraps existing multi-vector retrieval losses (such as CachedMultiVectorMultipleNegativesRankingLoss) and accepts a set of token_poolings parameters. None represents the original vectors, while HierarchicalTokenPooling(pool_factor=4) represents the compressed vectors. During training, Ward clustering uses a straight-through estimator to pass gradients from pooled vectors back to their member tokens without differentiating the cluster assignments themselves, allowing the encoder to learn from retrieval errors made with the compressed representations.

Aarsen trained two models on answerdotai/ModernBERT-base: one using only standard contrastive loss, and another using pooling-aware loss. Both were trained for one epoch on 120,000 MS MARCO triplets with seed 42, matching initialization and optimization settings exactly, differing only in the loss function. In the NanoBEIR@13 evaluation, the standard model retained 94.6% of its unpooled nDCG@10 at pool_factor=4 (49.03 vs. 51.84), while the pooling-aware model retained 97.8% (51.23 vs. 52.39).

Notably, the benefits of pooling-aware training extend beyond the pool_factor=4 used during training. Gains persist under fixed caps (reaching +3.48 percentage points at max_tokens=8 and +5.78 at max_tokens=4) and adaptive distance thresholds, despite training utilizing only full representations and proportional pooling conditions.

The applicability boundaries of this method warrant attention: all figures come from a single self-test by the author, the unpooled baseline itself was relatively low (51.84/52.39), the training scale was limited (one epoch, 120K triplets), and the evaluation set was NanoBEIR@13 rather than the complete BEIR benchmark. Aarsen himself noted that no hyperparameter search was conducted. The method has been integrated into the Sentence Transformers library, with code and model weights (tomaarsen/mve-ModernBERT-base-token-pooling) available for reproduction.

Sources :huggingface.co

Recherche

AWS Releases Universal Agent Evaluation Framework (UAEF) for Cross-Framework AI Agent Testing

Matériel
2026-11-11 23:50 GMT+8

Amazon Web Services (AWS) released the Universal Agent Evaluation Framework (UAEF) on November 11, 2026. This open-source system addresses the challenge of inconsistent evaluation standards caused by using different AI agent frameworks such as LangGraph, LangChain, Strands Agents, and Amazon Bedrock AgentCore. Previously, teams had to rebuild metrics and scoring logic for each framework's specific trace format, making it costly to switch tools or run agents on multiple platforms.

The core mechanism of UAEF involves adapters that convert execution records from various frameworks into a unified 'AgentTrace' canonical format. On this basis, the system scores agents using 36 built-in metrics across seven dimensions: tool calling, response quality, responsible AI, performance, multi-turn conversation, multi-agent coordination, and reasoning. All metrics are normalized to a [0, 1] interval, allowing agent versions from different frameworks to be compared on a single scale.

The framework also integrates personally identifiable information (PII) screening and stores results in Amazon DynamoDB and Amazon S3, enabling teams to save experiment data and generate reports. According to AWS, 12 organizations in the telecom, healthcare, and financial services industries have used UAEF to evaluate their AI agents, comparing versions before release to detect performance regressions.

UAEF is open source under the Apache 2.0 license on GitHub, providing a Python library, framework adapters, built-in metrics, and runnable notebooks. For developers building complex multi-agent systems, this tool offers a standardized evaluation infrastructure, helping to more objectively measure the impact of changes to prompts, models, or tools.

Sources :Amazon Web Services

Recherche

Genus's Gene Royalty Model: Controlling the Apex of the Genetic Pyramid and Building Barriers with Compound Data

Sujet · Genus基因特许权模式Matériel
2026-11-04

The business logic of UK breeding company Genus runs counter to intuition: it does not seek to raise more animals, but rather believes that fewer is better. According to Milena Mileva, investment manager at Baillie Gifford UK Growth Trust, in an on-site visit report written in November 2026, PIC (Pig Improvement Company), Genus's pig breeding subsidiary, operates core nucleus farms with high biosecurity isolation in Nashville, Tennessee. These farms house the apex of the global porcine genetic pyramid. Farmers access these elite genes by purchasing breeding stock or semen, while Genus's revenue comes from royalties collected as the genetics cascade through partners, multiplier farms, and large-scale producers.

The core mechanism of this model lies in data compounding. Mileva wrote: 'Genetics is a numbers game: more animals, more data and more generations give an R&D leader a compounding advantage.' This means that for every additional generation of data accumulated by first-movers, the cost for latecomers to catch up increases. Since its spin-off from the breeding and production arm of the UK Milk Marketing Board in 1994, Genus has expanded this compounding engine to both pigs and cattle through the 1999 acquisition of US bovine breeding company ABS and the 2005 acquisition of PIC, serving customers in over 80 countries.

PRRS-resistant pigs represent a critical step in taking this model from the laboratory to the marketplace. Porcine Reproductive and Respiratory Syndrome (PRRS) causes billions of dollars in productivity losses for the global pig industry annually. PIC spent over a decade developing a PRRS-resistant pig strain; last year, the US Food and Drug Administration (FDA) approved this gene edit, which Mileva called 'a major milestone towards commercialization.' At the Madison R&D center, she saw a red corporate flag signed by project team members, celebrating this regulatory breakthrough. However, commercialization is not yet complete: approvals in export countries are still underway, and customer willingness to pay remains to be verified.

China is viewed as Genus's greatest growth opportunity. Chinese consumers account for roughly half of global pork consumption, but the farming structure is transitioning from millions of smallholders to large, modernized farms. Mileva believes Beijing is supporting this modernization process; larger, professional farms are more likely to pay for long-term productivity gains, whereas smallholders are less concerned about genetic improvements that take several generations to yield returns. Genus's strategy is not to export live pigs to China, but to build a local genetic supply chain, collecting fees for the use of genetic technology through localized operations. The 2018 African swine fever outbreak caused approximately $100 billion in losses, exposing vulnerabilities in disease control under the smallholder model, but Mileva attributes the growing demand for genetics to policy-driven scale consolidation rather than the epidemic itself.

The boundaries of this argument warrant attention. Mileva acknowledges that the scale of the Chinese market is difficult to assess accurately—quoting PIC Chief Operating Officer Matt Culbertson saying 'nobody knows how many pigs there are in China.' The dividends of genetic improvement take several generations to materialize, contributing limited revenue in the short term. Breakthroughs in African swine fever resistance may not arrive until the 2030s. Furthermore, as an investment manager holding Genus stock, Mileva's visit reports naturally carry a portfolio perspective, and her judgments are not independently verified by third parties.

Sources :bailliegifford.com

Recherche

Google Releases A2UI 1.0 Open Protocol for Native Agentic Interfaces

Sujet · 谷歌A2UI协议Matériel
2026-11-11

US technology company Google released A2UI 1.0 on November 11, 2026, marking the first stable version of its open protocol for generative user interfaces. The protocol addresses the limitation of current AI agents relying primarily on text responses by allowing them to answer with interactive screens containing forms, charts, and other elements. Developers can now integrate using open-source renderer SDKs for Web, Android, and iOS, along with agent SDKs for Python, TypeScript, and Dart.

The core mechanism of A2UI separates "description" from "rendering." Agents describe what content needs to be shown via A2UI messages, while applications securely draw these contents using their own native components. This ensures that whether users access an agent within a company's own app or through third-party AI assistants supporting MCP Apps, such as Gemini or Claude, they receive low latency, high trust, and complete native fidelity. The architecture ensures sensitive data never reaches the model, and the UI streams in real-time as the agent works.

Version 1.0 introduces bidirectional function calls, allowing components like flight or reservation pickers to check availability instantly, and adds support for accessibility and extended metadata. Several major companies, including Atlassian, Alation, Bloomberg, and ByteDance, are already building products based on this protocol. For instance, Atlassian is bringing A2UI to Jira, enabling teams to interactively confirm details with AI agents directly within work items, while ByteDance's Lynx framework has released an A2UI renderer for cross-platform native display.

Sources :developers.googleblog.com

Recherche

Anthropic IPO Roadshow Discloses $17B Q3 Revenue and Mythos Agent's 3-Day AMD Port

Matériel
Vérifié 2026-11-12 00:39 GMT+8

US-based AI company Anthropic kicked off its Initial Public Offering (IPO) roadshow on November 10, disclosing key third-quarter financial metrics and compute strategy details on its official page. As of September 30, 2026, Anthropic reported quarterly revenue of $17.0 billion, representing over 12x year-over-year growth. However, the company posted a net loss of $23.3 billion, which includes a $21.9 billion non-cash charge related to changes in the fair value of convertible notes, derivative liabilities, and SAFEs due to increased valuation. The adjusted operating margin stood at 15%, while the actual operating margin was negative 3%.

Regarding compute infrastructure, co-founder Tom Brown outlined the concept of "total effective compute"—defined as physical capacity multiplied by discovered "compute multipliers"—in the roadshow video. He revealed that when Anthropic began working with AMD GPUs this year, it utilized an AI agent called Mythos to handle the porting process. According to Brown, Mythos enabled Claude to run on AMD hardware in just three days with virtually no human intervention, with Claude writing or porting nearly 100% of the kernels. This case serves as concrete evidence validating the feasibility of their multi-chip strategy, which had previously been considered costly due to the need for tailored software stacks for each chip type.

Additionally, Anthropic emphasized that its compute efficiency has improved by more than 4x since Q1 2025. These disclosures aim to demonstrate engineering capabilities and cost control mechanisms in heterogeneous computing environments to investors ahead of the planned Nasdaq listing under the ticker symbol "ANT".

Sources :ipo.anthropic.com

Recherche

Investigation: US Data Brokers Sell Half a Billion Dollars in RL Training Environments to Chinese AI Labs

Sujet · 美国强化学习环境对华销售Matériel
2026-11-11 21:51 GMT+8

According to a report by technology media outlet The Decoder, an investigation by ChinaTalk found that at least nine US data broker firms are selling approximately $500 million worth of AI training data to Chinese artificial intelligence laboratories. The data includes reinforcement learning (RL) environments—simulated worlds where AI models are trained on specific tasks such as coding, financial analysis, or office work.

The report highlights that Chinese buyers are deliberately purchasing the same products previously sold to frontier labs like OpenAI and Anthropic. By acquiring these pre-built environments, Chinese labs can skip costly in-house development and avoid wasted training runs. This strategy is particularly valuable given the acute memory bottlenecks affecting Chinese chips, allowing them to close the capability gap with less domestic compute.

In addition to purchasing environments, Chinese labs are reportedly using powerful US models like GPT and Claude as "teachers" through distillation, using their outputs as training examples. This two-pronged approach combines external data acquisition with knowledge transfer from US models.

The identified US firms include Mercor, Surge AI, AfterQuery, and Turing. The deals often run through offshore structures, with some procurement offices located in Singapore and San Jose. Because RL environments remain largely unregulated, existing export controls barely apply to this trade.

The authors of the investigation argue that domains with dual-use potential, such as coding and biology, pose particular risks. Coding environments can sharpen offensive cyber capabilities, while biology environments could enhance gene synthesis and bioweapons-relevant applications. They call for the US Bureau of Industry and Security to subject these dual-use environments to existing export controls.

Despite these concerns, the report notes that the relationships are not apocalyptic. Chinese labs still trail their US peers by at least a few months structurally. Furthermore, sensitive environments in areas like biology, chemistry, cybersecurity, and chip design are increasingly being built in-house by American frontier companies, limiting the scope of what is available for resale.

Sources :The Decoder

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

The clearer these constraints are, the clearer the scope within which an agent can advance autonomously. Human effort is thereby focused on goal setting, critical judgment, and risk control. The team described the expected evolution path as progressing from assisted execution, to autonomous completion of single-step tasks, and ultimately to completing a self-evolution loop within defined boundaries.

Evaluation infrastructure serves as the other pillar supporting this methodology. Around security investigation tasks, the team built SOCBench, which at the documented stage included 211 real-world investigation tasks covering 16 incident categories, evaluating five investigation stages across nine dimensions. The question set, execution configuration, and scoring specifications were all version-controlled, making results reproducible and releases comparable. Security investigations are long-horizon tasks; checking only the final conclusion misses many issues, so evaluation must also assess whether the environment was identified accurately, the entry point was confirmed, the main attack chain was reconstructed, the scope of impact was examined, and the final conclusion was supported by evidence.

A specific case demonstrated how evaluation transforms defects into regression assets. In one investigation, the report reversed the roles of attacker and victim. Root-cause analysis revealed that three modules lacked shared determination. The fix therefore expanded from correcting a passage in the report to a structured redesign: generating unified fields for the attacker, victim, and determination result, then having the report, timeline, and remediation recommendations reuse those fields. During retesting, the team reran the original scenario to confirm A was identified as the attacker and B as the victim, while verifying consistency across the outputs of the three modules. The role fields and retrieval logic subsequently entered the production release, and the corresponding regression case was added to the evaluation baseline. From then on, any change to the model, prompts, or rules had to use this sample to verify role determination and its cross-module consistency. One fix thus produced two outcomes: an effective product capability and a validation sample that could be run repeatedly.

Preliminary statistics show that the average cycle time from requirement creation to completion fell from 50.9 days to 13.8 days, while throughput within the measurement window rose from 63.9% to 98.9%. The team considers the next questions to validate as follows: Can the same mechanism be replicated from individual projects across more modules? Can practices dependent on individual initiative be institutionalized as stable team capabilities? Can shared outcome criteria become part of routine evaluation rather than working only during a particular project?