All Articles

An agent stops when its work looks done. Write the completion bar instead of asking for effort: a named check, required evidence, iterate until it passes.

xAI's Grok 4.6 keeps Grok 4.5's headline rates, but the $2/$6 tier only holds below a 200K-token prompt: cross it and the whole request reprices.

xAI's launch table sets Grok 4.6's default effort against GPT-5.6 Sol's and Fable 5's documented ceilings. Why the rung changes what a score means.

Both terminal harnesses ship bash sandboxing off by default. The differences that hold up are openness, model freedom and what each vendor publishes.

Background agents get more done unattended and hand you a large diff to review cold. Four criteria decide which mode a given task actually belongs in.

Four levers decide what a vendor benchmark table means: which benchmark version, which task subset, what shape the metric takes, and who got a column.

A buyer-side checklist for giving an agent standing access to business systems: what to scope, who holds the secrets, and how revocation works.

Moving from prompting an agent to employing a persistent teammate changes identity, memory, standing tool access and who owns the unwatched outcome.

Demonstration-captured routines are quicker to create but harder to diff, version and audit than built workflows. Where each paradigm actually fits.

xAI's Grok Bot launched in beta on August 11. The launch page says Bots have their own computer; xAI's own docs say an account's Bots all share one.

Two of the three plans that unlock Grok Bot's beta are Cursor tiers, and per xAI's own docs its authentication and data governance run through Cursor.

Meta released Muse Glimmer 30B under a genuine Apache 2.0 license: a dense, multimodal model distilled from Muse Spark and aimed at local agent work.

NVIDIA's 30B, 3B-active Nemotron 3.5 Lightning trails Qwen 3.6 35B A3B on 13 of 14 rows in NVIDIA's own table. The pitch is speed, not peak accuracy.

Sakana's sovereign-AI strategy is not pretraining: Namazu is a Kimi K2.6 fine-tune, and Sakana validated Fugu's conductor on a Gemma 4 base days later.

Remotion makes video a React build target, but an agent cannot easily judge whether the finished motion is any good — and the licence turns on company size.

A triage for the first jobs you hand an AI teammate — recurring, bounded, reversible, verifiable — plus the work that should stay with a human.

Claude Code documents a --worktree flag, Copilot CLI's /worktree is experimental, Muse Code's subagent isolation is reported. Why agents need it, and its cost.

Vercel's skill packs bundle agent skills behind one install command and an update command, with no versioning, checksums or provenance documented.

Anthropic flips Claude Code to auto mode by default on Pro, Max and Team plans from August 14. What changes, what to pin, and what the studies show.

Claude Code 2.1.224 adds cross-session SendMessage, self-hosted runners on Team and Enterprise, and drops the 200-subagent cap. What ships, and what to guard.

Codex CLI 0.147.0 imports Cursor-managed skills and syncs imported Claude and Cursor conversations. What is actually portable across agent harnesses now.

GitHub's enterprise managed settings now allowlist MCP servers, and the Copilot impact dashboard adds an ROI section GitHub itself calls directional only.

Google is extending Limited Ad Serving to all Google Ads, rolling out gradually and completing by 2028. Approved status no longer means your ads will serve.

Mistral's Shieldstral reads your safety policy as prompt text at inference time. A 3B Apache-2.0 guard model with a mixed, not one-sided, benchmark record.