aidigest — 2026-07-23

/code-review now runs as a background subagent — Review work no longer floods your main conversation with its intermediate reasoning; it runs off to the side and returns findings. In the same window Claude also stopped auto-running /verify and /code-review on its own — you now invoke them explicitly. Net effect: cleaner context and no surprise review passes eating tokens mid-task. If you relied on the auto-behavior, wire it into a Stop hook instead. source

/subtask for in-session subagents; /fork goes background — /subtask launches subagents inside your current session, and /fork was repurposed to copy the whole conversation into a new background session while you keep working. This is the manual lever for the fan-out-and-keep-only-conclusions pattern Igor already leans on — worth adding to muscle memory next to the automatic Agent dispatch. source

Session-wide caps on WebSearch, subagents, and long MCP calls — There's now a default ceiling of 200 WebSearch calls and 200 subagent spawns per session, and any MCP tool call running past two minutes auto-backgrounds. For long headless/cron runs (like this scout), these are guardrails against runaway loops — but if you build a genuinely wide fan-out job, know the ceiling exists and is configurable rather than getting silently truncated. source

Worktree subagent isolation bug fixed — Subagents launched with isolation: 'worktree' were running git-mutating commands against the main repo instead of their isolated copy — exactly the failure mode worktrees exist to prevent. Fixed now. Also fixed: the ultracode keyword opt-in was firing on non-human input like webhook payloads, which matters if you pipe untrusted text into headless sessions. source

Claude Code now ships as Bun-compiled-to-Rust — Since ~v2.1.181 the CLI is Bun built in Rust rather than the old JS build (Simon Willison reverse-engineered the binary to confirm). Jarred Sumner's line: "Startup got 10% faster on Linux but otherwise, barely anyone noticed." Boring on purpose — the point is a more stable, self-contained binary running on millions of devices, not new features. Good reminder that the harness itself is now a compiled artifact you can reason about. source

"Claude Code sends 33k tokens before reading your prompt" — A widely-shared teardown (705 pts on HN) measured Claude Code's system-prompt + tool-schema overhead at ~33k tokens vs OpenCode's ~7k on an empty session. Whether or not you switch harnesses, the practical takeaway is real: the base context tax is large, so aggressive tool deferral, lean skill sets, and compaction matter more than people assume — especially on long autonomous runs where that overhead recurs. source

clawk — give agents a disposable Linux VM, not your laptop — 226 pts. Runs coding agents inside a throwaway VM with its own kernel: host filesystem was never mounted, outbound traffic is allow-listed (DNS resolves but unlisted TCP/UDP/ICMP is refused), your ssh-agent is forwarded so git push works without keys entering the VM. The payoff is running agents with no permission prompts because the boundary contains them. Directly relevant to Igor running headless agents on a VPS full of other services. source

deja-vu — searchable memory over your existing agent sessions, synced by SSH — 131 pts. Retroactively indexes the session transcripts your agents already wrote to disk (searches 3.3 GB in ~12 ms), redacts credentials at index time, and syncs append-only batches between machines with plain ssh/scp — idempotent, no server. This is the pragmatic, boring version of "agent memory": mine what you already have instead of a new cloud dependency. Close cousin to the profile-mining this very repo does. source

Rewriting Bun from Zig to Rust with agents — the $165k receipt — Simon Willison's writeup of Jarred Sumner using Claude agents for a large-scale language-runtime port, with roughly $165,000 in token spend (5.9B input, 690M output, 72B cached reads). The striking part isn't the cost — it's the role shift: instead of reviewing 1M+ changed lines by hand, Sumner supervised the generation process, reading outputs for issues and re-prompting Claude to fix the loop rather than fixing code directly. Adversarial review plus a strong test suite as the feedback signal. That's the industrial-scale agent supervision pattern. source

danluu: testing-heavy, review-light loops beat human review at agent scale — danluu's agentic-coding notes (179 pts) argue that when agents produce hundreds of changes a day, human review can't keep up, so lean on tests instead — but "write tests" isn't enough: point agents at fuzzing and randomized testing on risky code, and have independent agents repeatedly re-verify each alleged bug repro to crush false positives. This is the verification-loop discipline, made concrete. Worth stealing for the trading/quant pipelines. source

Simon Willison: delegate implementation to cheaper subagent models — "Judgement" makes the case for letting the main model design/review/synthesize while handing the grunt implementation to subagents on lower-tier models — implementation work rarely needs the top model, and you burn premium tokens far slower. The Workflow/Agent tools already expose per-agent model and effort overrides; this is the mental model for when to reach for them. source

Does codebase cleanliness change how well agents perform? — A controlled minimal-pair study (210 pts) finds that cleaner code doesn't change whether the agent finishes the task — but it cuts token consumption by 7-8% and file re-visits by 34%. So the payoff of readable code isn't success rate, it's efficiency: the agent thrashes less and burns fewer tokens. Concrete grounding for "invest in readability so agents run cheaper" when structuring repos agents will live in. source

Added — subtasks-fork (in-session subagents via /subtask and background-session copies via /fork), agent-sandbox (run agents in a disposable VM / allow-listed sandbox before granting broad autonomy).

Updated — code-review now notes it runs as a background subagent and is no longer auto-invoked (call it explicitly or wire a Stop hook); subagents notes the new per-session spawn cap and cheaper-model delegation; headless-sdk notes the Bun-in-Rust compiled binary.

Removed — none.