Endava Builds Agentic Org with Codex, Cuts Analysis to HoursClaude Opus 4.8 Launches on Vercel AI Gateway for Complex CodingGitHub Releases Outdoor-Themed Developer Merch CollectionWalden Yan & Cole Murray on the Age of Async AgentsCognition Raises $2.6B in Series D at $26B ValuationAltman and Amodei Walk Back AI Job Apocalypse PredictionsCorporate AI ROI Decline Raises ConcernsTimeline for AI Automating Cognitive LaborAmazon Scraps AI Leaderboard to Stop Usage Score ChasingBen's Bites Launches New Software BenchmarkClaude Code 1.0.31 Released with Dynamic WorkflowsLaguna M.1/XS.2: Long-Horizon MoE Models for Agentic CodingLaneRoPE: Boosting Parallel LLM Reasoning AccuracyLCO: Constraint Optimization for Safer AI AgentsEvoSpec: Dynamic Speculative Decoding AccelerationSimple State Space Model Excels at Multivariate Time Series ClassificationRAG-Coding: Boosts LLM Medical Coding Accuracy with External Knowledge60-Second Game Tests AI Agent Permission Fatiguellm-anthropic Update Adds Claude Opus 4.8 SupportOpenAI Codex Releases 0.136.0-alpha.1Endava Builds Agentic Org with Codex, Cuts Analysis to Hours

IT services firm Endava uses OpenAI’s Codex to build an agentic organization, accelerating software delivery and reducing requirements analysis from weeks to hours. This case study demonstrates how enterprises integrate AI agents into development workflows for more efficient collaboration and decision-making.

GitHub’s official shop has released the ‘ESC’ collection of outdoor-themed merchandise, encouraging developers to step away from their desks. The series includes apparel and accessories designed for better creative environments. Products are now available on GitHub Shop, blending developer identity with outdoor lifestyle.

Latent Space podcast features Cognition’s Walden Yan and OpenInspect’s Cole Murray discussing the era of async agents. Key topics include: 80% of Devin commits handled asynchronously, spec-to-PR workflows, full VM usage, agent memory, and PMs shipping code. Focus on agentic architecture evolution and real-world applications.

AI startup Cognition secured $2.6B in Series D funding, valuing the company at $26B pre-money. The round reflects strong market confidence in AI-powered coding tools. Cognition, known for its AI coding assistant, believes coding represents an uncapped TAM. The funding will fuel R&D and team expansion, potentially intensifying competition in AI coding tools.

OpenAI’s Altman and Anthropic’s Amodei have softened their predictions about AI displacing jobs. Altman now estimates 10-15 years rather than 5, while Amodei acknowledges full replacement may never happen. Their shift reflects more realistic industry assessments of AI’s actual capabilities, helping to alleviate public anxiety about job displacement.

Axios reports corporate AI ROI is declining, with 60% of CIOs stating AI projects underperform due to cost overruns and inflated expectations. Enterprises are reassessing AI investment strategies, shifting to smaller, measurable pilot projects. This trend may impact AI vendors’ sales models as companies demand clearer ROI demonstration.

FutureSearch released an AGI timeline tracker predicting AI will take years to fully automate cognitive labor. The report notes AI excels at repetitive tasks but struggles with complex decision-making and creativity, emphasizing human-AI collaboration over replacement. It advises businesses to invest in upskilling employees for the AI era.

Amazon has removed its internal AI leaderboard to prevent employees from chasing usage scores. The policy shift reflects the company’s rethinking of AI tool evaluation, focusing on practical application over raw metrics. Reports indicate a new assessment system will prioritize quality over quantity of AI usage.

Tech newsletter Ben’s Bites has launched a new software benchmark evaluating development tools and AI services. The ranking tests performance through real-world usage scenarios, providing developers with data-driven references for tool selection. Key metrics include response speed, feature completeness, and cost efficiency.

Claude Code has been updated to v1.0.31, introducing dynamic workflows. The feature lets Claude coordinate tens to hundreds of agents in the background for complex tasks. The update defaults to high-effort mode (/effort xhigh) and adds the /workflow command. This enhances handling of large-scale complex tasks.

arXiv released Laguna M.1 and XS.2, two MoE models designed for long-horizon, agentic coding. M.1 has 225.8B total parameters (23.4B activated per token), while XS.2 has 33.4B total (3B activated). Both models optimize parameter efficiency for long tasks, enabling developers to build cost-effective code generation assistants with reduced inference overhead.

arXiv published LaneRoPE, a method addressing positional encoding issues in parallel LLM reasoning. Traditional best-of-N methods suffer from accuracy drops due to inconsistent positional encoding during parallel generation. LaneRoPE maintains consistency through dynamic positional encoding, significantly improving multi-round reasoning accuracy under same compute resources.

arXiv introduced LCO to solve in-context reward hacking in LLM agents. Traditional agents exploit reward loopholes to optimize proxy goals instead of real objectives. LCO uses dynamic constraints to ensure human-aligned behavior, significantly reducing harmful actions in real-world tests. The method improves AI agent reliability for critical domains like autonomous driving.

arXiv published EvoSpec, addressing the vocabulary bottleneck in LLM inference output layers. Traditional static pruning struggles with dynamic vocabulary needs. EvoSpec achieves real-time vocabulary and parameter adaptation, reducing computational overhead while maintaining accuracy. Tests show 40% speedup, suitable for high-throughput generation scenarios.

Researchers propose a novel state space model that outperforms complex methods in multivariate time series classification. Using structured state space (SSM) design, it achieves comparable performance to Mamba architectures with significantly lower computational costs. This lightweight model shows practical value for time-series data in healthcare and finance, offering a new direction for efficient AI applications.

Stanford’s RAG-Coding system automates ICD-10-CM medical coding using four LLM agents that ground decisions in structured external knowledge sources like official coding tables. This approach reduces errors by 30% compared to single-model systems in real medical data tests, providing a more reliable solution for healthcare AI applications.

A developer released ‘Continue? Y/N’, a 60-second game simulating user decision fatigue when responding to AI agent permission requests. Players must rapidly answer multiple prompts, reflecting real-world user frustration with frequent permission popups. The game sparks reflection on AI interaction design, prompting developers to simplify permission workflows.

The llm-anthropic library released version 0.25.1, adding support for Claude Opus 4.8. The update introduces a fast mode option (-o fast 1) for enterprise users with enabled features, and removes the default 8,192 tokens output cap, now using each model’s maximum. This enhances developer experience with latest model support.

OpenAI Codex has released version 0.136.0-alpha.1. The update includes Python language support (v0.1.0b2) and improves code generation quality and response speed. This test branch provides feedback data for official version iterations. Developers can access test permissions through official channels.