When to Build an Agent Skill, When to Use a Prompt, and What You’re Running When You Install One
When a skill actually beats a prompt, an AGENTS.md entry, or an MCP server, and the security cost of installing an executable folder from a 40,000-skill marketplace.
Google's new Gemini coworker agent gets its own Workspace account, email, and calendar. The real decision is whether you can govern it like the non-human identity it is.
Amazon Blocked AI Shopping Agents the Same Week Meta Built a Standard to Let Them In
Amazon blocked outside shopping agents the same week Meta and Sierra published an open standard to let them in. The real question underneath is an access-control one.
Claude Haiku 5.5 is 90% cheaper and leaps on computer use, but it matched GPT-6 Luna's price rather than undercutting it, and the benchmark it leads with runs at max effort. Where the budget model...
Reflection AI's Beam is a 501B open-weight frontier model pitched against DeepSeek. The weights, the price, and a clear benchmark win aren't here yet. Where it actually fits.
You Can Prove What Your AI Agent Cost. Proving What It Did Takes a Different Kind of Log.
Token dashboards measure the model. The agent fails at the trajectory level, where wrong tool calls and retry storms never hit a cost chart. How to trace an agent's steps, run evals on the...
