aidigest — 2026-08-16

Weekly scan of the Claude Code / agent-harness world, filtered for what a builder like Igor can actually use. Two threads this week. First, the multi-session story from last week keeps maturing: you can now @-mention another running session by name, subagents fork with the parent's conversation and prompt cache for free, and Anthropic published a plain how-to on keeping sessions cheap. Second, the ecosystem is in a "small, fast, single-binary agent" phase — Bullet, Hax, Ante, Mole all shipped in one week — while two grounding reads (danluu on language choice, Cactus's 14 MB on-device agent) push back on the hype in useful directions. A practical week, not a hype one.

What changed in Claude Code

Mention another session with @; subagents now fork by default (v2.1.232) — Type @ in a prompt to reference another live Claude Code session by name, and SendMessage now delivers to a bare name when exactly one session matches — so last week's cross-session messaging is now something you drive by name, not by hunting IDs. Bigger under the hood: subagent forking is on by default, and a forked subagent inherits the parent's full conversation and prompt cache. For Igor's parallel-worktree and orchestration runs that means a spun-up subagent starts already knowing the context and costs far less to launch. source

Todo/task-tracking tools removed from Opus 4.8, Sonnet 5, and newer models (v2.1.233) — The harness no longer hands the built-in TODO/task-tracking tools to the current-gen models; they're expected to manage multi-step work themselves rather than lean on a scaffolded checklist tool. It's a small line in the changelog but a real signal about where the harness is going: less rigid tool scaffolding, more trust in the model's own planning. If a workflow of yours depended on seeing an explicit todo list emitted, that behavior changes on these models. source

High-effort /code-review now runs as a background agent (v2.1.232) — When you run /code-review at high effort levels it now executes as a background subagent instead of blocking the session, so a deep review doesn't freeze your main thread or flood it with findings. Pairs naturally with the "run it before you ship" habit — you can kick off a thorough review and keep working while it churns. source

The containment fixes keep coming (v2.1.232–233) — Another batch of permission-boundary holes closed: a PowerShell bypass via variable-writing parameters, a Windows bypass using Cygwin-style symlinks (then partly reverted in 233 for ordinary symlinks — worth noting if you're on Windows), nested git repositories wrongly inheriting a parent directory's trust, and Windows NT \??\ device-prefix paths dodging UNC validation. Bash input redirections are now permission-checked like argument spellings. Same lesson as last week: the week autonomy is the default is exactly when to keep updating and not assume last month's sandbox still holds. source

Hot in agent land

Anthropic's own guide to keeping Claude Code sessions cheap (306 pts) — A rare plain-language post from Anthropic on the mechanics of session cost, not the marketing. The reusable rules: /clear between unrelated tasks so old context isn't resent; /compact before a break because prompt caches expire after an hour and summarizing while still cached is far cheaper; /context in a fresh session to spot dead weight (e.g. an MCP server you can disable); @-mention files instead of typing paths so they attach directly and skip a Read call; set /model and /effort at the start since changing them mid-conversation blows the cache; and use /rewind instead of /compact for minor cleanup because earlier turns stay cached and it's free. Directly relevant to a cron scout that pays per token. source

A wave of small, single-binary coding agents (Bullet, Hax, Ante, Mole) — Four terminal agents hit HN's front page in a week, all optimizing for lean and fast: Bullet (113 pts) routes easy work to fast models and escalates only when needed, uses targeted search instead of embedding the whole repo (~8% of context in its demo), and claims 95.8% on SWE-Bench Verified; Hax (115 pts) is a minimalist terminal-native agent in C; Ante (167 pts) runs offline as a single binary; Mole (96 pts) is a deep-research agent for the terminal. The common thread — model routing, retrieval over whole-repo embedding, tiny footprints — is worth borrowing even if Igor stays on Claude Code. source

"What's the best programming language for coding agents?" — mostly a myth (261 pts) — danluu tests the popular claim that dynamic languages cost 1/2–1/3 the tokens of Rust/Go/C++ for LLM coding. On real tasks (a zstd decoder, a Pandoc-style converter) rather than toy benchmarks, that gap largely vanishes — at higher effort, static languages match or beat dynamic ones, and if anything static typing's feedback loop helps at scale. The stronger signal is language popularity: mainstream languages get more-correct, cheaper solutions; obscure ones (J) underperform despite theoretical token efficiency. For Igor's Python/Rust stack the takeaway is freeing — pick for the team, not for a token myth. source

New approaches worth knowing

Needle2: a 14 MB agentic model for wearables, smart home, and robots (529 pts) — Cactus shipped an Apache-2.0, 45M-parameter model quantized to a single 14 MB binary that runs in ~28 MB of RAM — on an ESP32/STM32 microcontroller, a Raspberry Pi (500+ tok/s), a budget phone, or a Quest. It's not a chatbot; it's built for three things — tool calling with typed parameters, device control (smart home, robots), and structured extraction — with confidence scoring that escalates uncertain requests to the cloud. For Igor's home-automation interest this is the concrete shape of a private, always-on, offline anticipatory assistant: a tiny local model that maps intent to device actions and only phones home when unsure. source

Codex generated a full game in 52 minutes with aggressive sub-agent use — Simon Willison logged Codex (on GPT-5.6 Sol Ultra) building a complete playable game, "Moonlight & Mayhem," in 52 minutes by leaning hard on parallel sub-agents. It's a demo, not a study, but it's a useful datapoint on where the frontier of one-shot autonomous generation sits right now — and a reminder that the sub-agent fan-out pattern Claude Code exposes is exactly what makes these long-horizon runs work. Contrast with last week's two-weeks-and-unfinished desktop rewrite: bounded, buildable scope is where autonomy shines. source

Playbook changes

Added — context-cost-hygiene: the concrete cost levers from Anthropic's session guide (/clear, /compact, /context, /rewind, @-mentioning files, and setting /model+/effort up front to avoid cache invalidation).

Updated — cross-session-messaging now notes @-mentioning a session by name and SendMessage delivery to bare names (v2.1.232); subtasks-fork notes subagent forking is on by default with full conversation + prompt-cache inheritance, making /fork and in-session forks cheap; code-review notes high-effort runs now execute as a background agent; auto-mode updated to past tense — it became the default for Pro/Max/Team on 2026-08-14 and is now live.