Researched and written with 313,408 tokens (in 10 · out 29,043 · cache-read 66,585 · cache-create 217,770) across 5 Claude calls

Coding agents: model releases and prompting

• Quoting Prompting Claude Fable 5.2: Anthropic's guide says Fable 5.2 follows what the prompt says about code ownership, so framing the codebase as yours changes how bold its edits are.
• Simon Willison on the same tip: telling Fable 5.2 "This codebase is yours" makes it less cautious when improving existing code, so read the updated prompting guide on every new model.
• llm-anthropic 0.31: adds Fable 5.2 and Mythos 5.2 support to the llm CLI; Fable 5.2 currently ranks first on Artificial Analysis.
• Code-only benchmark thread: the gap between models shrinks on code-only benches, and xhigh/max effort score worse than high on code, so don't default to max effort for coding.
• Opus 5.5 vs Codex in T3 Code: Theo charts Opus 5.5 overwhelmingly dominating Codex usage over time in T3 Code, which is a usage signal, not a benchmark.

Plan limits and cost

• Ultra plan value thread: the $200 Ultra plan yields about $15,000 of Opus 5.5 but only about $3,000 of Fable 5.5, with Fable capped at 50% of the weekly limit, so Opus is the better value for bulk work.
• 1.1M request analysis: the numbers behind the plan-value claim come from 1.1M real Anthropic API requests on an Ultra sub, so they carry real weight.
• Amazon replaces Sonnet 4.6 with Strands Decider: a ~98% cost cut on ticket labeling; commenters note it is really a return to cheap classifiers, and the "millions" only holds at about 92M tickets a year.

Agent harnesses and control surfaces

• pingdotgg/t3code: an "agent harness control surface" with mobile, web and desktop apps that drives your local Claude Code, Codex, Cursor, OpenCode and others via their existing subscriptions.
• T3 Code remote control overhaul: cmd+k "mcp" sets up a connector so any agent can control T3 Code from another device.
• T3 Connect MCP auth: the MCP link now authorizes through your T3 Connect account instead of copy-pasted pairing codes, which makes pairing environments far easier.
• T3Code takeover from a laptop: a user reports taking over remote environments from a MacBook through a built-in bot-style experience.
• T3 Code "tab complete" prompts: experimental toggle for tab-complete prompt suggestions in the harness.
• Codex rust-v0.169.0-alpha.8: another alpha of the Rust Codex CLI; the release page captured no useful notes.
• Claude Code v2.1.322: a new Claude Code release, but the captured page shows no changelog, so check the notes directly.

Skills, memory and context engineering

• mattpocock/skills: small, composable, model-agnostic agent skills for real engineering, pitched against process-owning frameworks like GSD, BMAD and Spec-Kit.
• garrytan/gbrain: self-hosted, MIT-licensed shared memory for your agents, with sourced facts, corrections and withdrawal, starting from keyword retrieval and adding semantic search later.
• Agentic AI Engineer 2026 checklist: argues for learning context engineering and strict tool schemas, because agents otherwise re-read files every loop or forget earlier constraints.
• Compaction/200k fix in Chat/Cowork: a Reddit thread on working around the 200k compaction limit in Chat/Cowork, which Code doesn't have.

CLI/TUI

• MD Context Manager with Embedded CLI (Kanban Pro): a markdown context manager with a built-in CLI for Claude workflows; the post has no body text, so treat it as low-signal until you see a demo.