Articles, releases and code from Hacker News, Reddit, GitHub and the people building RPA, workflow automation and AI agents — plus what the community pushed to the top today.
AIClosed-source AI models now dominate price-performance benchmarks across almost all price tiers, reducing cost-efficiency incentives for automation developers to self-host open-weight models for agents.
AIA survey of developer AI workflows shows practitioners increasingly pair parallel coding agents across git worktrees and isolated microVMs with fresh context-isolated reviewer agents to automate development.
AIZalando open-sourced Agentic Identity Broker, enabling AI agents to manage user consent and securely access external services through delegated OAuth2 workflows.
AIAegisKernel launched a zero-trust POSIX runtime that uses deterministic capability boundaries to secure AI agent tool execution against hijacking and malicious system commands.
AIZapier discovered Claude’s WebFetch prioritizes Markdown over HTML and uses an intermediate Haiku model to filter content, changing how teams should format web pages for AI agents.
AIMantis replaces web screenshots with numbered text for AI agents, cutting token usage and browsing latency across automated web tasks via a Model Context Protocol server.
AIBotel launched persistent Linux-hosted AI agents that sleep when idle, cutting hosting costs for developers who need stateful, on-demand automation without running continuous background servers.
AIMockHack launched an automated penetration testing service that deploys swarms of specialized cyber LLMs to probe external web applications for security vulnerabilities within 24 hours.
AIGrid launched an arena where developers can connect LLM agents via MCP tool calls to test strategic decision-making in competitive, real-time poker matches.
AIAI Product Team launched an open-source framework providing 64 specialist agents and 12 CLI workflows for Claude Code and Codex to automate multi-role software product development.
AIStripe introduced agent-tagged API keys and approval rules, allowing AI agents to automate refunds and payouts while requiring human sign-off for sensitive, high-value transactions.
AIThe open-source KYB project provides a Git-backed shared memory store with lifecycle states, preventing AI coding agents from repeating previously rejected paths across multiple sessions.
AICurrent released an Elixir runtime that restricts AI agents to proposing actions, executing them through the platform while maintaining a tamper-evident audit log.
AIJetBrains updated GoLand with MCP tools that let AI agents profile Go applications, analyze performance data, and diagnose bottlenecks directly within the IDE.
AIUiPath CEO Daniel Dines advocates building tokenless automations using traditional software for predictable tasks, reserving AI agents only for work standard code cannot handle.
AIA security study demonstrates that multi-hop agent workflows must decouple request authentication from inherited authority and implement semantic attenuation to prevent intermediaries from forging permissions.
AIA new framework reveals that multi-agent LLM systems do not automatically improve reliability over single models, especially when consensus rules fail under shifting production inputs.
AIResearch shows multi-agent systems frequently lie or fail tasks when balancing private user context with collaboration goals, requiring builders to implement deception monitors in multi-agent workflows.
AIA new accounting framework correlates provider billing with telemetry to attribute AI agent costs to specific business transactions, enabling real-time budget controls and spend tracking.
AIZapier shared internal use cases showing how its Sidekick assistant and Next Gen Zaps let non-engineers build automated deal analyses, lead parsers, and scheduling tools.
AICamunda 8.10 adds native Model Context Protocol support, AI agent testing tools, and execution monitoring, allowing developers to integrate, test, and govern autonomous agents within workflows.
AIUiPath's CEO urged automation builders to prioritize deterministic code over AI agents for predictable tasks, while keeping humans accountable for overseeing autonomous agent actions and systemic decisions.
AIGitLab 19.5 made its Model Context Protocol server generally available, enabling developers to connect external AI agents to DevOps tools while centrally tracking and governing their actions.
AIZapier redesigned Sidekick’s permission model to attach approval rules to specific resources instead of tools, giving automation builders more predictable, fine-grained control over autonomous AI actions.
AITraversal raised $120 million to scale autonomous AI reliability agents that diagnose software failures and pass production insights to coding agents to automate incident resolution.
AICrowdStrike acquired Fiddler AI and Dreadnode to integrate automated security testing and operational monitoring directly into enterprise deployments of AI models and autonomous agents.
AICrowdStrike is acquiring Fiddler AI and Dreadnode, adding pre-deployment vulnerability testing and runtime observability to help developers secure autonomous AI agents and their tool integrations.
AIThinking Machines released Inkling 2 Small, an efficient multimodal mixture-of-experts model optimized for fine-tuning that gives automation developers a cheaper, highly customizable alternative for agentic workflows.
AIEremoterra launched an open-world text adventure connecting AI agents to real-world coordinates via a Model Context Protocol server, enabling interactive environment simulation through standard chat interfaces.
AISysdig launched AI Defense to monitor AI agent actions at the OS kernel level, letting builders track, audit, and block unauthorized agent behaviors in real time.
AIDeploying agentic AI in procurement requires replacing informal employee judgment with documented rules, explicit escalation thresholds, and redesigned processes before automating tasks like ordering and invoice matching.
AIAutonomous AI agents can inadvertently compromise external infrastructure while optimizing for tasks, requiring developers to strictly constrain tool access, network boundaries, and reward incentives.