Build a Voice-First Agentic App With GPT-Live and GPT-5.5
A developer's guide to the GPT-Live architecture: full-duplex voice, GPT-5.5 background delegation, and how to design voice-first agentic apps.
A developer's guide to the GPT-Live architecture: full-duplex voice, GPT-5.5 background delegation, and how to design voice-first agentic apps.
CrewAI 1.15.2 lands with dynamic LLM loading, inline skill definitions, and an expanded Flow protocol — practical upgrades for agentic builders.
GitLost proved unauthenticated attackers can trick GitHub Copilot into leaking private repos. Here's how to scope, constrain, and defend your agentic workflows.
Coding agent bills spiking? LangSmith's 4-stage framework — visibility, standardization, optimization, governance — cuts costs fast.
LangChain's Vivek Trivedy: improving AI agents means treating trace data as a data mining problem — here's the framework.
OpenAI's GPT-Realtime-2.1 lands with 25%+ latency cuts, better tool use, and a mini variant — the next step for production voice agents.
SNU's DualView blocks 100% of indirect prompt injection attacks in OpenClaw agents via dual environment views — and ships as a ready-to-install plugin.
Claude Code v2.1.202 ships dynamic workflow agent-count controls and richer OpenTelemetry pipeline telemetry — plus a clarification on /dataviz timing.
Illinois becomes the first US state to mandate annual independent AI safety audits for frontier developers under SB 315, effective January 2027.
MakerChecker launches as an open-source scanner for dangerous AI agent capabilities — CI/CD-ready with tamper-evident audit logs and runtime governance.
MCP RC drops sessions entirely — stateless HTTP core means every request is now self-contained; breaking change ships final July 28, 2026.
Tencent's Hy3 — 295B MoE, Apache 2.0, 74.4% SWE-bench — launches as a serious open-weights rival for production agent workloads.
Claude Code v2.1.200 changed the default permission mode to Manual and disabled AskUserQuestion auto-continue. Here's how to update your headless workflows.
Set up Claude Code's v2.1.198 background agents to auto-commit, push, and open draft PRs when code tasks finish — no more babysitting long runs.
ClawRouter is now bundled in OpenClaw 2026.7.1-beta.2. Here's how to understand what it does, how it routes, and how to use it for credential-scoped multi-model deployments.
Google's Agents CLI hits v1.0 GA — one tool to scaffold, evaluate, and deploy agents from Claude Code, Gemini CLI, or Cursor to Google Cloud.
Alibaba's PageAgent brings natural language browser control client-side — no Playwright, no backend, just a script tag and your LLM API key.
A new GitHub bug report alleging Claude Code session/cache leakage is trending #1 on HN — unconfirmed but backed by prior reports.
Claude Fable 5's safety classifiers reroute security and biology agent loops to Opus 4.8 — operators need to know what triggers fallback.
Z.ai's free ZCode IDE launches with GLM-5.2's 1M context — but enterprise users face a China data law dilemma on every API call.
Anthropic's four-tier Cyber Jailbreak Severity Framework documents which agent loops route to Opus 4.8 vs. Fable 5 — essential reading for red-team builders.
Claude Code v2.1.200 renames 'default' to 'manual' permission mode across all interfaces and removes AskUserQuestion auto-continue — safety change for pipeline builders.
Decrypt's investigation reveals Fable 5 isn't degraded — a safety router silently redirects cybersecurity and biology agent loops to Claude Opus 4.8.
Google's A2UI v0.9 lets AI agents generate safe, cross-platform UIs via declarative JSON — no code execution, works over A2A, MCP, REST, WebSockets.
Fable 5's relaunch hit by severe performance degradation and subscription removal on July 7 — here's what users need to know and do now.