Latent Space
·
1 day ago
Superapp usage grew by 1 million users since the previous report, while OpenAI's Codex and ChatGPT Work products saw usage jump 2.5x in a week. PrismML released Bonsai 27B, a compressed model that runs on consumer devices at 3.9 GB and 1.125 effective bits while supporting multimodal and agentic workflows. The shift toward local inference and compressed models is making edge deployment viable for serious agent applications, while industry focus is moving from raw model quality to harness quality and observability as primary differentiators.
Latent Space
·
1 day ago
Developers are shifting focus from autonomous agents themselves to the systems that manage them, with "loop engineering" becoming central to how AI agents are controlled and evaluated in production. A survey by Barr Yaron identified coding agents as a key trend, with tools like Claude Code, Cursor, and Gemini CLI replacing traditional IDE-based development by handling broader objectives across multiple files. Enterprises are now deploying AI engineering specialists called forward-deployed engineers to implement AI capabilities, with companies building "software factories" that automate parts of the development lifecycle while keeping humans in control of critical decisions.
MarkTechPost
·
1 day ago
Mistral Vibe for Code, Claude Code, OpenAI Codex, and Cursor were scored on five dimensions (scaffolding, testing, PR workflows, surface coverage, and cost/control) across a common engineering task of adding a subscriptions endpoint with tests and opening a pull request. Mistral Vibe for Code scored 22/25, leading on cost at $14.99/month, open-source availability, and self-hosting capabilities, while Claude Code and Codex tied at 21/25 but at higher costs ($20–$200/month). The comparison used documented features and published benchmarks as of July 14, 2026, rather than executing agents directly, and explicitly noted that benchmark figures across different suites are not directly comparable.
The New Stack
·
1 day ago
Tailscale is hosting a webinar on July 28, 2026 addressing how enterprises can manage network access for both human users and AI agents using a unified architecture rather than separate tools. The event will cover how traditional VPN and privileged access management tools designed for humans fall short when dozens or hundreds of AI agents need secure access to corporate networks. Organizations need consistent access policies that can handle requests from any source—developers, contractors, pipelines, and AI agents—while maintaining security audits and the ability to revoke access when tasks complete.
AWS Machine Learning
·
1 day ago
Thrad.ai built a multi-agent system using Strands Agents and Amazon Bedrock to automate lead research across six social platforms (Hacker News, Reddit, Stack Overflow, GitHub, dev.to, ProductHunt), reducing manual research time from 30-45 minutes per lead to an automated pipeline. The system deployed two orchestration patterns—Swarm and Graph—with Graph achieving 32-second average latency and $0.06 cost per prospect versus Swarm's 45 seconds and $0.08 cost, though Swarm produced higher email quality (8.2 vs 7.6 on human rating). Organizations can now automatically discover buying signals across multiple sources, score prospects 0-100 using weighted criteria and temporal decay, and generate personalized outreach at scale.
AWS Machine Learning
·
2 days ago
Amazon extended QA Studio, an agentic QA automation tool built with Amazon Nova Act, to support batch regression testing through test suites that execute use cases in parallel on AWS Fargate, and added a command-line interface for CI/CD pipeline integration with OAuth authentication and environment variable overrides. Test suites can run 20 concurrent tests instead of sequentially, reducing overall execution time while organizing tests by functional area or testing purpose. The CLI enables automated quality assurance in CI/CD pipelines with three exit codes distinguishing test failures from infrastructure errors, allowing teams to integrate AI-powered visual navigation testing into deployment gates.
NVIDIA
·
2 days ago
NVIDIA argues that performance per watt is the critical efficiency metric for AI infrastructure, as it determines how many tokens can be generated within a fixed power budget. The NVIDIA Blackwell NVL72 platform delivers up to 25x better performance per watt compared to the Hopper generation when serving mixture-of-experts models across a 72-GPU domain. This efficiency advantage will determine which organizations can scale their AI operations in a power-constrained environment.
Menlo Ventures
·
2 days ago
Enterprises face a shift from human-led cybersecurity to AI-native defense architectures as attackers increasingly use machine-speed automation, requiring a three-layer system of behavioral context engines, autonomous response agents, and continuous validation. The behavioral engine learns what normal identity behavior looks like to detect anomalies, with Abnormal AI reporting detection accuracy orders of magnitude faster than human review and serving over 25% of Fortune 500 companies. Organizations that adopt these AI-native defenses will compound advantages weekly, while those delaying face an increasingly irreversible gap as the window to rebuild closes.
The New Stack
·
2 days ago
Port, an agentic software development company, launched its Port AI Builder service to address concerns about ungoverned AI-assisted coding by adding governance controls, human review, and organizational context to code generation. The platform uses a Plan Mode that drafts code, asks clarifying questions, and requires human approval before building, with all plans versioned and audited for traceability. The service aims to shift software quality assurance from manual processes to AI validation combined with human oversight, enabling teams to build production-grade code without accumulating technical debt from rushed development.
The New Stack
·
2 days ago
Model Context Protocol (MCP) faces criticism from developers for adding unnecessary complexity in small projects, but the skeptics overlook that enterprise agentic systems require MCP's capabilities for credential delegation, audit trails, and structural least privilege that direct API calls cannot provide. MCP enables security teams to embed authorization context at the tool-call level and structurally prevent agents from exceeding their assigned scope, whereas allowlists alone are insufficient governance in regulated environments. Enterprise adoption of agentic systems will require solving two remaining challenges: making capability-scoped server provisioning accessible to non-specialists and providing operational visibility tools for security teams to manage MCP connections across their environment.
The New Stack
·
2 days ago
Thira, a startup founded by Apptio's Sunny Gupta and Kurt Shintaffer, is building AI agents to automate IT back-office processes that CIOs currently handle manually across multiple systems. The company raised $21 million in seed funding led by Madrona and has signed ten design partners ahead of a fall launch, initially targeting two repetitive IT processes out of six or seven identified. Thira differentiates itself through trust features like audit trails and rollbacks, offering semi-autonomous and fully autonomous modes, and plans to integrate with existing enterprise platforms rather than replace them.
The Neuron
·
2 days ago
Visa built an agentic AI workflow within Elastic Security to automate triage of high-stakes mainframe security alerts previously requiring manual investigation. Triage time for the mainframe identity detection dropped from 10–20 minutes to seconds by chaining two queries, adding an LLM validation step, and delivering results directly to the incident response ticketing system via webhook. The four-stage pipeline pattern is now reusable across other detections and may absorb additional response automation currently handled by a separate SOAR platform.
The Neuron
·
2 days ago
AgentKey launched a unified API platform that consolidates access to search, social media, finance, and cryptocurrency data sources through a single authentication key for AI agents. The service operates on a subscription model with shared credit balances across all integrated services, priced at rates like $3.55 per 120 web searches and $1.40 per 450 Twitter data requests. Users can now connect multiple data sources in one installation step without managing separate API keys, vendor dashboards, or per-service billing.
The Neuron
·
2 days ago
Osaurus is a macOS application that runs open-source AI models locally on Apple Silicon Macs without sending data to external servers. The software is free, MIT-licensed, and allows users to integrate ChatGPT, Claude, or Gemini for specific tasks while maintaining a shared memory across all models. Users can build autonomous agents with voice control, file monitoring, and browser extensions that operate offline and independently.
The Neuron
·
2 days ago
Fabraix Playground is a testing platform that lets users attempt prompt injection attacks on AI agents to discover security vulnerabilities. The tool provides an interactive environment where researchers can practice exploiting these agents before malicious actors find the same weaknesses. Organizations can use the results to patch vulnerabilities and improve their AI systems' defenses against prompt-based attacks.
Apple ML Research
·
2 days ago
Researchers introduced Pare, a framework that simulates active users to evaluate proactive AI agents in digital environments by modeling apps as finite state machines rather than flat APIs. The framework includes Pare-Bench, a benchmark with 143 tasks across communication, productivity, scheduling, and lifestyle applications that test context observation, goal inference, intervention timing, and multi-app coordination. This approach addresses limitations in existing proactive agent evaluation methods that fail to capture the stateful nature of real digital interactions.
