Researched and written with 315,493 tokens (in 10 · out 35,598 · cache-read 66,228 · cache-create 213,657) across 5 Claude calls

Coding agents

• mattpocock/sandcastle: TypeScript lib to run coding agents in isolated sandboxes via one sandcastle.run() call, with branch strategy and commits merged back; provider-agnostic with native OS sandboxes (Bubblewrap on Linux/WSL2).
• Putting in a ton of work into Sandcastle v1: v1 pitch is a "software factory" with 6 built-in harnesses (Claude Code, Codex, etc.), Daytona/Vercel/E2B sandbox integrations, and Inngest/Trigger.dev templates for durable runs.
• Been trying out DeepSeek harness: reports high cache hit rate and good UX, but no editing of previous messages.
• Maybe Redux, or a Redux-like framework needs a comeback: single-source state restricts the agent's decision space, so one way of doing things suits agents.
• I got tired of AI agents saying they were done when the browser had actually failed: addresses agents falsely claiming completion after browser failures, a core verification gap in agent loops.
• Automating /clear session: approach to automating Claude Code /clear, relevant to context hygiene.
• Fable 5.1 vs Opus 5.5: community comparison of the two latest Claude models for coding work.

LLM mechanics: tokens, caching, context

• I benchmarked 10 open-source AI memory systems on BEAM (100K) with Claude: two systems matched full-context accuracy while reading 24-35x less text, so memory layers can cut context cost without quality loss.
• 5 ways to see token usage in Claude Code and what each one counts: compares token-usage views in Claude Code and what each measures, useful for trusting your cost numbers.

Harness wars: T3 Code vs pi

• Still hard for me to believe that T3 Code is one of the most engaged repos on all of Github: signals T3 Code has very large community traction.
• T3 Code now supports managed policies: org-level policy config for IT teams, a push into enterprise.
• RT @davis7: This was a good setup ~3 months ago: davis7 has moved off pi to T3 Code, citing its better built-in orchestrator and GUI-first workflow.
• and there is one who is actually building the best one: counter-take praising davis7's pi harness for controlling many agents.

CLI/TUI

• ghostty-org/ghostty: fast, GPU-accelerated, platform-native terminal emulator, also embeddable as a library via libghostty.
• Show HN: OmniDiff: Rust content-aware git difftool using content-specific algorithms; commenters expect the AGPL-plus-donations license to stall adoption.

Agentic SaaS and tools

• agentmash: open-source multiplayer AI for teams' coding agents.
• Termy: coding agent surfaced in the Mac notch.
• Click: managed agents API that works with any model.
• Riverside Agents: team of AI agents for video production, editing, clipping and posting.
• I built a shared knowledge store for agents with Claude Code: free-to-try Claude connector for a shared agent knowledge store.
• Claude Code made my app's 54-second 4K60 launch film: Claude Code produced a launch video from real screen recordings, a creative-automation example.