Researched and written with 308,645 tokens (in 10 · out 31,455 · cache-read 66,075 · cache-create 211,105) across 5 Claude calls

Models, limits & cost

• Claude Opus 5.5 is the first model in the Claude 5.5 family. Anthropic says it matches Fable 5.1 on most tasks at 40% lower cost than Opus 5, which makes it the default to test for agent workloads.
• Agents SDK drains 5h limits faster: Theo confirmed a bug on Anthropic's side where all Agents SDK usage burns the 5-hour limit faster than Claude Code does. Weekly limits are unaffected, and a fix is pending.
• HiRoute (site) is a local, open-source (Apache-2.0) router for Codex and Claude Code. It scores each message's complexity and sends easy work to cheap models and hard work to the main model. It claims about 90% cost cuts at equal quality and says it surfaces failures instead of silently downgrading.
• Discarded code patterns burning API budget is a Reddit thread arguing that tokens spent on code patterns the agent later throws away can account for half your spend. No body text came through, so the claim is from the title only.
• Unified Claude GA for Enterprise merges chat and Cowork and adds four admin cost controls, including an "Efficient mode".

Coding agents & harness

• Claude Code Projects now admits every Pro and Max user from the waitlist.
• Gemini CLI is no longer optimized for the latest Gemini models. Unpaid-tier and Google One users were moved to Antigravity CLI on June 18, 2026, and Gemini 4 Argon requires Antigravity.
• Agent Toolkit for AWS is an official set of AWS-supported MCP servers, skills and plugins. It works with Claude Code, Codex, Cursor, Devin and Kiro, so you can have agents deploy to AWS with guardrails.
• T3 Code "share with guest" is a beta feature that lets you send a link to a thread so a coworker can add context directly. It replaces the copy-paste and screenshot loop.
• T3 Code daily improvements: Theo says he will ship one meaningful improvement to T3 Code every day, matching Codex's "28 days of improvements" pace.

Context, search & session tooling

• Graft parses a repo into a symbol and call graph that agents can query (map, signatures, callers, references, ranked files). It uses tree-sitter with no model or API key, and stores the graph as project files instead of in a vector DB.
• AgentsView imports session histories from Codex, Claude Code, Cursor and 20+ other agents into local SQLite. You can search messages and tool calls and get cross-agent token and cost reports.
• Persistent memory system for AI agents is a Reddit post from an author asking whether anyone wants it. It is useful mainly as a gauge of demand and prior art for agent memory.

Running agents unattended

• Swipe: Agents that work while you sleep covers tools to keep agents running after the laptop lid closes, keep API keys out of the chat, and hand work between parallel sessions.
• AgentValley is a Mac notch app that shows the status of every running Claude Code agent at a glance.
• Remarc is a Mac menu-bar tool for attaching a text selection, screen capture or web element to a comment. A connected agent reads the comment over MCP and marks it in progress or resolved, which helps when a visual bug is hard to describe.
• Swipe: an agent where your mouse is features AtAt, where typing @@ in any Mac text field runs an agent inline.
• Swipe: Deployment made simple covers a CLI with agent skills for deploying to a VPS from your coding agent, plus workflows for feeding it an X account and old sessions.

Skills & design guidance

• mattpocock/skills is a small, composable, model-agnostic set of agent skills from Matt Pocock's .agents directory. It is pitched as an alternative to process-owning frameworks like GSD, BMAD and Spec-Kit.
• Nutlope/hallmark is an anti-AI-slop design skill for Claude Code, Codex, Cursor and OpenCode. It picks three themes from 48, builds the page around the visitor's goal, and lints the result.
• pbakaus/impeccable is one design skill with 24 commands, live browser iteration, and 62 deterministic detector rules for AI-generated frontend patterns.
• skill-creator is installed with npx skills add anthropics/skills --skill skill-creator. It drafts a skill from your workflow and tests it on real prompts, which suits tasks you keep re-explaining to the agent.
• claude-howto is a visual, example-driven guide to Claude Code covering agents, hooks, skills and MCP, with copy-paste templates.

CLI tooling

• Homebrew 6.0.0 adds tap trust, a faster and smaller default JSON API, Linux sandboxing, and brew bundle improvements. Commenters report that it force-upgrades casks marked auto_updates with no pinning or cooldown, and some are moving to mise and MacPorts.