Researched and written with 315,472 tokens (in 10 · out 34,858 · cache-read 66,315 · cache-create 214,289) across 5 Claude calls

Coding agents & harnesses

• Claude Code v2.1.312: adds oidc.user_id_claim for the apps gateway (e.g. Entra oid) and click-to-send on queued messages in fullscreen.
• Codex 0.165.0: built-in Google Cloud Vertex AI provider (model_provider = "gcp") and GPT-6.1 Terra added to the model catalogs.
• How Anthropic ships Claude with Claude: inside look at Anthropic's improvement and investment loops, framed by the S-1 numbers and the Opus 5.5 launch.
• How Anthropic ships the #1 LLM safely: Claude Platform engineers explain separating model "brain" from agent "hands" for reliable managed agents, which is the key architecture idea for harness design.
• Harness Engineering from Scratch: 21K-star open curriculum (14 lectures, 8 projects) on building the engineering environment around a coding agent.
• Matt Pocock on agent-friendly tasks: tasks where the agent has all the data, can experiment, verify independently and hill-climb are where trust and results peak; useful framing for choosing what to delegate.
• Coding From the Himalayas With an Offline AI Agent: practical setup for running a coding agent with no reliable network or power.
• GitHub Copilot server-to-server auth: GitHub Apps installed on orgs can now call Copilot SDK-built services and agents without a user token (public preview).

LLM mechanics & model comparisons

• Context Engineering (DeepLearning.AI x Chroma): free course on feeding agents the right information, format and timing, since model quality degrades as context grows.
• Kimi K3.1 Pro vs GPT-6 Astra: same four build tasks produced near-identical results, with Kimi costing $21 against $54, a cost data point for model routing.

CLI/TUI

• Ghostty 1.4: big performance and memory gains, the first non-Kitty terminal with a complete Kitty graphics implementation, and better SSH integration; commenters still miss iTerm2's tmux -CC on Linux.

Agent orchestration & software factories

• SuperPlane 1.0: open-source "software factory" that scans the backlog and coordinates agent sessions and CI, with the team reporting 52% of merged PRs automated last month.
• Corvid: free, source-available dashboard for running many Claude Code agents at once.
• Theo's T3 Code fleet: 26 concurrent threads across machines with one-click "update all" for T3 Code, Codex and Claude Code, relying on thread restore after update.

Agentic SaaS (Product Hunt)

• PulseDev: click a UI bug on localhost and your AI agent fixes it.
• Revix AI: AI code review that learns a repo's own rules.
• Paddock: computers and sandboxes for agents.