aidigest — 2026-08-02

Weekly scan of the Claude Code / agent-harness world, filtered for what a builder like Igor can actually use. Claude Code itself had a quiet, stabilization week — the big features (Opus 5, nested subagents, sandbox switches) all landed in the previous drop, and this week is bug-fixes. The energy moved to the ecosystem, where two currents run in opposite directions: builders are racing to run many agents in parallel (multiplayer harnesses, tmux orchestrators, local merge queues), while a sober counter-current — a long-horizon benchmark and a formal-verification stunt — insists you still can't let them run lights-off.

What changed in Claude Code

A stabilization week: v2.1.220 is bug-fixes only — After the feature-heavy v2.1.219 (Opus 5 as default, nested subagents to depth 3, sandbox.network.strictAllowlist, workflowSizeGuideline), the latest release is reliability work with no new surface. Nothing to adopt, nothing to re-learn — a good week to let your setup settle rather than chase knobs. If you pinned a model id or a workflow-size setting last week, it's still valid. source

New DirectoryAdded hook — v2.1.219 quietly added a hook that fires after /add-dir (or the SDK's register_repo_root) registers a new working directory mid-session. That's a real new automation point: if Igor wants a fresh checkout to auto-run setup, re-index, or announce itself when it's pulled into a session, this is where to wire it — deterministic, so the model can't forget. Worth folding into the hooks toolbox alongside SessionStart/Stop. source

Headless hardening that matters for this very scout — Several v2.1.219/218 fixes target non-interactive runs: claude -p no longer drops the answer it already produced when a turn dies on a mid-stream API error; the stream-json init event now lists mcp_server_errors for --mcp-config entries that failed validation; and nested subagent text now forwards through stream-json with --forward-subagent-text. If you run agents from cron or the SDK (this digest does), these close exactly the silent-failure gaps that make headless runs hard to trust. source

Hot in agent land

qm — a "multiplayer" agent harness (650 pts) — A YC-backed open harness built around the idea that a whole team shares agents, not just one person. A headless TypeScript/Fastify core drives multiple backends (Claude Code, Codex, OpenCode, Pi), keeps per-user sessions and memory in Postgres, and — the interesting part for security — has each agent act as the person using it with that person's real credentials, every action audited, rather than handing agents broad standing permissions. This is the "agent-as-user, gateway-mediated secrets" pattern (cf. last week's OneCLI) hardening into a full platform. source

Opus 5 on SlopCodeBench: still can't run lights-off (405 pts) — A humanlayer writeup benchmarking Opus 5 on a long-horizon eval that reveals requirements in checkpoints instead of all at once — closer to real work. Opus 5 leads (24% strict pass vs 6% for Opus 4.8 / Sonnet 5) but degrades over a session: after clearing early checkpoints it accumulated defects in every later one, wrote 5× more functions and ~29k lines (51% tests), and 93% of its output tripped "slop" flags. The author's line: until models clear ~80%, "today's models can't be relied on to run lights-off without steering." A useful counterweight to the parallel-fleet hype below. source

agent-manager — a tmux TUI for a fleet of coding agents (97 pts) — Runs Claude Code, Codex, OpenCode, Grok and Gemini CLI side by side, each in its own tmux session so they survive you quitting the manager. Live status per session, a project tree, prompt-injection without attaching, optional per-session git worktrees, and a Ctrl+R diff reviewer where line comments flow back into the agent's conversation. If Igor is already fanning out work, this is the "cockpit" layer for watching several agents at once without tab-juggling — no framework, just tmux. source

New approaches worth knowing

Invert code review: trust a 93-line spec, not 1000 lines of AI code (115 pts) — A formally-verified 3D mesh-intersection project where the AI wrote the implementation (~1000 lines) and its Lean proofs (~60k lines) with zero human review, and the human trusts only two things: a 93-line formal specification and the Lean checker that proves the code satisfies it. The generalizable move — don't hunt for bugs in AI logic, pin down what correct means and let a checker enforce it — is the strongest version of the verification-loop idea for anything safety- or money-critical (relevant to Igor's quant/trading infra, where a tight invariant beats a code read). source

A local merge queue for parallel Claude Code agents (42 pts) — The unglamorous plumbing the parallel-fleet dream actually needs. When several agents work in separate worktrees, letting them all push creates races, redundant heavy builds, and shared-resource test flakiness. This tool makes each agent run land instead of pushing: rebase onto integration, wait in a machine-local FIFO, run your test command, then push only when no other lane is mid-push. Zero infra, no billed CI minutes, no PRs. A concrete answer to "how do I actually run five agents without them stepping on each other." source

MarbleOS: the agent GUI as a workspace, not a chat log (131 pts) — The most-discussed "what should agent UIs look like?" post this week argues the chat thread is the wrong primitive — it buries the files, tools, tasks and outputs an agent is actually touching. Its pitch is a dashboard-style workspace where all of that is visible at once. Still early/beta, but the direction is worth tracking: as you run more autonomous and parallel work, scrolling a transcript stops scaling, and "show me what the agent is doing" becomes the real UX problem — the same instinct behind agent-manager's diff pane and this scout's static issue pages. source

Playbook changes

Added — none. No genuinely new coachable Claude Code features shipped this week (v2.1.220 is bug-fixes); the week's notable tools (qm, agent-manager, merge-queue) are third-party and don't belong in a Claude-Code-native playbook.

Updated — hooks now lists the new DirectoryAdded event (fires after /add-dir or SDK register_repo_root) among its example triggers.