aidigest — 2026-07-26

Opus 5 is the new default — claude-opus-5 replaces 4.8 as the default Opus model: 1M context, and fast mode priced at $10/$50 per Mtok. Opus 4.7 was dropped from fast mode (/fast now applies to Opus 5 and 4.8), and the bundled claude-api skill now defaults to Opus 5 with a documented migration path from 4.8. For Igor: anything that pins a model id in scripts or the budget tooling should be re-checked, and the 1M window changes what "too big to fit" means for the trading/data pipelines. source

Nested subagents are back — up to depth 3 — A week ago v2.1.217 turned off nested subagents by default and capped concurrent subagents at 20; v2.1.219 reverses course and lets subagents spawn their own subagents up to depth 3 by default. Net: real orchestration trees (a lead agent fanning out to teams that themselves fan out) are supported again, but with a sane concurrency ceiling so a runaway fan-out can't melt the box. Worth knowing the exact depth if you design multi-tier workflows. source

Built-in sandbox gets network + filesystem switches — New sandbox.network.strictAllowlist denies any non-allowlisted host for sandboxed commands, and sandbox.filesystem.disabled (v2.1.216) lets you skip filesystem isolation when you don't want it. This is Claude Code growing the same containment story the clawk/disposable-VM crowd has been pushing — but native, so you can allow-list egress for a headless agent without standing up a separate VM. Directly useful for running agents on a VPS that also holds secrets and other services. source

Dynamic workflows now have a size guideline — A new workflowSizeGuideline setting plus a default "medium" cap (fewer than ~15 agents) gives the model advisory guidance on how big a dynamically-authored Workflow should get, and the current default shows up in the running-workflow status line. If you were letting Workflow auto-scale fan-out based on a token budget, this is the lever that keeps it honest instead of silently ballooning. source

Fewer surprise auto-runs; forked skills go background — Continuing last week's "stop auto-invoking things" trend, /deep-research now only starts when you manually invoke it, and skills declared context: fork now run in the background by default rather than blocking the main thread. Predictable token spend and a cleaner main conversation — the same discipline this scout depends on for long headless runs. source

Anthropic: the new rules of context engineering for Claude 5 models — An official post (278 pts on HN) arguing that the older, guardrail-heavy playbook actively hurts newer models. Concrete moves: delete overconstraining rules in favor of judgment ("match the surrounding style"), drop usage examples from tool descriptions, push verification/review guidance into progressively-disclosed skills rather than the system prompt, and keep CLAUDE.md lean — Anthropic says it cut 80% of Claude Code's own system prompt with no performance loss. If Igor's CLAUDE.md files have accreted rules, this is permission to prune them. source

"Why Software Factories Fail" — harness engineering isn't enough — A widely-shared essay (384 pts) making the counterweight case to pure automation: loop/harness tuning speeds up individual builds but can't fix a training-level gap — models are rewarded for passing tests, not for maintainability, so architectural rot doesn't surface until incidents weeks later. It cites a Faros report of a ~243% jump in incidents per PR post-AI-adoption, and recommends a human-in-the-loop, design-first flow (product → architecture → program design → vertical slices) targeting a safe 2–3× gain rather than a fantasy 10–100×. Sober reading for anyone tempted by lights-off agents. source

OneCLI — keep real secrets out of the agent entirely — 106 pts. A credential gateway: your agent gets a placeholder like FAKE_KEY, and when its HTTP request passes through the gateway, OneCLI matches host/path patterns and swaps in the real (AES-256-GCM-encrypted, centrally vaulted) key before forwarding. So a compromised or prompt-injected agent can't exfiltrate the actual secret, and you get central rotation plus an audit trail of which agent hit which service. A cleaner answer than .env-in-the-worktree for headless agents that need live API keys. source

Fireside chat with the Claude Code team (Cat & Thariq) — Simon Willison's writeup of a conversation with two of the builders. Highlights: the system prompt shrank ~80% with newer models ("removing examples was extremely helpful, because it was just more creative than the examples we gave it"); they now let Claude do full code review on non-critical areas after months of eval work identifying which files never regressed, with humans gatekeeping core systems; and heavy internal use of auto mode is what gave them confidence to ship it, having "mitigated every single issue" red-teamers found. The meta-lesson: as implementation gets cheap, the bottleneck moves to product judgment. source

Simon Willison: cheap reverse-engineering flips the home-automation ROI — The argument that coding agents don't just make automation faster — they change the economics. The old blocker for automating an undocumented home device wasn't only the upfront reverse-engineering effort but the psychological weight of maintaining brittle glue code forever. When an agent can rebuild it in an afternoon, "throw it away and start again" carries far less baggage, so projects that were never worth it suddenly are. Relevant to Igor's smart-home / anticipatory-assistant tinkering: the bar for "worth automating" just dropped. source

Added — auto-mode (dynamic-permission autonomous runs, the feature Anthropic leans on internally), deep-research (/deep-research multi-source research, now manual-invoke only), and model-fast-mode (Opus 5 as default + /fast and /model for cost/latency tuning).

Updated — subagents now notes nested subagents to depth 3 and the 20-concurrent cap; workflows notes the workflowSizeGuideline / default-medium sizing; agent-sandbox now covers Claude Code's built-in sandbox.network.strictAllowlist / sandbox.filesystem.disabled alongside external VMs.

Removed — none.