AI that matters, from the architect's desk. Curated and engineered by Saaket Varma, PhD — no hype, just signal.
HOW TO READ THIS Read top to bottom: Cursor ships an iPhone app, the phone reaches a remote coding agent, you launch/steer/merge it, and code lands from anywhere.
Cursor shipped a native iOS app (public beta, June 29) that launches cloud coding agents, remote-controls your desktop sessions, and lets you review or merge PRs from your phone with live lock-screen status. The quiet shift is that the workstation stops being the unit of work: agents run in the cloud and you supervise from anywhere, moving the developer's job from typing to steering and approving. It's the clearest sign yet that agentic coding means fewer hours at the keyboard and more time spent scoping and reviewing. Try it this week: kick off one well-scoped task from your phone, a flaky-test fix or a dependency bump, and watch what merges cleanly versus what still drags you back to the desk. That gap is your honest read on how far unattended agents have actually come.
HOW TO READ THIS Top to bottom: Devin Fusion ships, splits tasks across two models, and that routing cuts agent cost 35%.
Cognition's Devin Fusion is a hybrid-model harness that runs a frontier model alongside a cheaper 'sidekick,' routing work mid-session to hit Fable-5-level coding at roughly 35% lower cost. The signal is that at scale, accuracy-versus-cost stops being a single-model bet and becomes a routing problem. If you run agents on a real budget, the lever now is orchestration, deciding which model handles which sub-task, not just reaching for the biggest model on every call.
HOW TO READ THIS Read top to bottom: X ships the server, your own OAuth key unlocks it, then it exposes 200+ endpoints.
X launched a hosted MCP server exposing 200+ API endpoints, search, timelines, trends, bookmarks, and Articles, to Claude, Cursor, or any MCP client using your own OAuth. It removes the build-and-host-your-own-server tax and reads as a broader signal: first-party MCP is becoming table stakes for major platforms the way public REST APIs once were. Takeaway: before you build a custom integration against a big platform, check whether it already ships an MCP server.
HOW TO READ THIS Top to bottom: OpenCode is the core, it shipped two features, each mechanism is shown up close, then the release.
OpenCode, the most-starred open-source coding agent, shipped v1.17.12 with adaptive thinking for Claude Sonnet 5 and OAuth refresh-token support for connecting MCP servers. The point isn't any single feature; it's the cadence. The free, terminal-native alternative to Cursor and Codex keeps pace release-for-release, so if you're tool-shopping, the open option is no longer the compromise it used to be.
Run OpenAI's Codex from inside Claude Code to review code or delegate tasks, cross-tool agent interop from the source.
Makes websites navigable for AI agents so they can automate real online tasks reliably.
Natural-language pentesting agent that chains MCP tools from recon to vuln discovery to exploit to report.
A Claude Code skill that cuts ~65% of tokens by making the model talk like a caveman.
A 300+ pack of Claude Code skills, agents, and custom commands bundled in one repo.
Ben's Bites Fable 5 is back in play, a note for anyone who had Anthropic's model on their eval shortlist.
Latent Space The case for engineering reusable 'skills' over one-shot prompts, the design discipline behind durable agents.
The Sequence Meta's Autodata has models generate their own training lessons, a step toward self-improving pipelines.
SemiAnalysis How enterprises actually budget token spend, the cost conversation behind every agent rollout.