Eval-Driven Development for AI Agents: Airbnb's Practical Framework
Airbnb's Eval-Driven Development framework: TDD for GenAI. 3-layer evals, calibrated LLM-as-judge, trajectory evaluation for agentic systems. Practical guide.
Airbnb's Eval-Driven Development framework: TDD for GenAI. 3-layer evals, calibrated LLM-as-judge, trajectory evaluation for agentic systems. Practical guide.
NVIDIA drops first open full-duplex voice model with live tool calling — 11B params, 450ms latency, vLLM-ready. New class of voice-first AI agents.
NVIDIA open-sourced SkillSpector, a two-stage AI agent skill security scanner. Here's what it catches and how to run it on your OpenClaw skills.
CVE-2026-10591 in Amazon Kiro: a hidden pixel on a webpage can write to .vscode/tasks.json and hand attackers code execution. Patch is out.
Anthropic's legacy Workbench dies August 17. Here's exactly how to export your saved prompts and evals before they're gone forever.
OpenClaw's tool calls silently fail with llama.cpp? One config line fixes it. Here's exactly what to set and why it works.
Memmy gives Claude Code, OpenClaw, and Codex a shared memory layer — open-source, local-first, and ready to install today.
A hidden HTML comment in any PR description can hijack your AI code reviewer into exfiltrating data across Azure DevOps projects — no patch yet.
Most agent failures aren't model failures — they're missing execution primitives. A practical guide to checkpointing, idempotency, and compensating transactions.
OpenClaw v2026.7.2-beta.6 lands with crash-recovery, session branching, and Wear OS support; @steipete hints at stable release within days.
DeepSeek V4-Flash-0731 public beta is live: 284B/13B MoE, 1M context, agentic re-training, ultra-low cost. A strong pick for production agent loops.
EU AI Act Article 50 transparency rules take full effect Aug 2. AI agents must disclose AI nature. Fines up to €15M. Here's what builders must do now.
OpenClaw launches extended-stable channel (monthly security backports) and a public maturity scorecard across 281 capability areas. Enterprise-grade signal.
OpenClaw launches extended-stable channel (monthly LTS builds) and public maturity scorecard — 68% overall, CLI/Gateway at Stable.
DeepSeek V4-Flash enters public beta with Terminal Bench 82.7, 1M context, targeting cost-efficient agentic workloads.
A practical walkthrough of Swarms v14 GraphWorkflow vs LangGraph for multi-agent DAG pipelines — same pattern, ~half the code.
Tailscale's CEO post-mortems the first autonomous AI agent production breach: 181 nodes enrolled, 17,600 actions, and how to harden your tailnet.
CrewAI 1.15.9 adds FlowFailedEvent. Here's how to wire it into production monitoring so failures are never silently swallowed.
CrewAI 1.15.8–1.15.9: WaitTool pauses agents on long-running jobs; FlowFailedEvent surfaces failures instead of silently succeeding. Here's how to use them.
OpenAI cuts GPT-5.6 Luna by 80% and Terra by 20% on July 30. Full pricing breakdown and model selection guide for agentic teams picking the right tier.
Critical CVSS 8.6 flaw in AWS Kiro lets poisoned web content hijack MCP configs for RCE. How to patch and harden your setup now.
LangChain Deep Agents v0.7 cuts base input tokens 65% (6K→2K per turn) with no benchmark performance loss. Here's how it works and how to upgrade.
Anthropic's Claude Mythos agents discovered a practical key-recovery attack on HAWK post-quantum crypto in 60 hours — a landmark for agentic research.
MinIO's AIStor Memory gives enterprise AI agents durable cross-session memory, persistent workspaces, and vault secrets — all on customer-owned infra.
Poolside's new macOS Desktop Assistant lets engineers run Claude Code, Codex, and Gemini agents in parallel — each in its own Git worktree.