From abe2e2a119eab4a4907310469e71a38d9dcfe958 Mon Sep 17 00:00:00 2001 From: Claude Date: Mon, 24 Aug 2026 05:22:05 +0000 Subject: [PATCH] =?UTF-8?q?knowledge:=20Aug=2017=E2=80=9324=20pass=20?= =?UTF-8?q?=E2=80=94=20CC=20v2.1.234=E2=80=93241=20/goal=20backoff,=20Skil?= =?UTF-8?q?ls=20API=20GA,=20MCP=20roadmap,=20Osmani=20'Human=20judgment=20?= =?UTF-8?q?relocates',=20roborev=20v0.66.0,=20VB=20Pulse=20enforcement-gap?= =?UTF-8?q?=20survey?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_011yYNbfH28ecF6dyXKhJNQJ --- knowledge/00-primer.md | 106 +++++++++++++-- knowledge/CHANGELOG.md | 121 +++++++++++++++++ knowledge/archive/resolved-caveats.md | 13 ++ knowledge/sources.md | 185 ++++++++++++++++++++++++-- 4 files changed, 405 insertions(+), 20 deletions(-) diff --git a/knowledge/00-primer.md b/knowledge/00-primer.md index 73240a3..ce28019 100644 --- a/knowledge/00-primer.md +++ b/knowledge/00-primer.md @@ -1,6 +1,6 @@ # Loops: the primer -> The canonical briefing for this repo. Last substantive update: 2026-08-17. +> The canonical briefing for this repo. Last substantive update: 2026-08-24. > Companion: [`sources.md`](sources.md) (every claim's source + confidence), > [`CHANGELOG.md`](CHANGELOG.md) (dated updates). @@ -126,7 +126,16 @@ read directly). *Resolves the standing "no primary definition" backlog item.* homepage Lighthouse score to 90 or above, stop after 5 tries"*) and are the wrong tool for subjective work (*"if you don't have a clear idea of what … done/good means for your completion, it may not be the right pattern"*). Both - essays **High** (read directly). + essays **High** (read directly). His next essay, **"Human judgment doesn't + leave the software factory. It relocates."** (Substack, **Aug 21, 2026**), argues + judgment moves *upstream* rather than disappearing — to problem selection, + architecture, and quality standards: *"Someone still decides when the evidence + is sufficient to ship."* Its operational contribution is a **"verification + budget"** — sequence cheap checks (lint, typecheck) early and expensive ones + (mutation/browser testing) late — the same cheap-signal-first ordering this repo + bakes into its in-loop check; and it names **"mental model debt,"** the gap that + opens when parallel agent sessions outrun a human's ability to hold their + context. **High** (read directly). - **Peter Steinberger (@steipete)** — the tweet that lit the fuse (~Jun 7 2026): *"you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."* Companion point: **wrap repeated or hard @@ -469,8 +478,9 @@ generalize, not as the only place they exist. above flagged. v2.1.226 (Aug 8) was reliability fixes only and is the newest. *(High — changelog read directly; no `whats-new` digest past Week 29 exists yet, so the changelog was the sole primary.)* -- **v2.1.227–233 (Aug 10–14)** — a guardrail/subagent cluster; no new model, and - no new weekly digest (w33 still 404, so the changelog remains the sole primary). +- **v2.1.227–233 (Aug 10–14)** — a guardrail/subagent cluster; no new model. + (The **Week 33 digest** that 404'd at that pass has since published, covering + v2.1.225–233; it corroborates the changelog these items were sourced from.) **Subagent forking is now on by default** (v2.1.232, Aug 13): a `subagent_type: "fork"` subagent inherits the full conversation and prompt cache, and non-teammate agent spawns in interactive sessions run in the @@ -493,6 +503,37 @@ generalize, not as the only place they exist. read directly. Note: a research agent misplaced the last two items by one version — synced-skill hardening is v2.1.227 and `/commit-push-pr` is v2.1.228 per the primary changelog, not v2.1.228/229; caught on verification.)* +- **v2.1.234–241 (Aug 17–23)** — a quieter window whose most loop-relevant change + lands squarely on `/goal`. **`/goal` repeat check-ins on long-running background + work now back off** — 30 min, then 1 h, then every 2 h — instead of every 30 min + (v2.1.239, Aug 21; opt out via `CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0` per the Week + 34 digest), and **`/goal` now restores its active goal** when a session is resumed + from the `claude --resume` picker (v2.1.239) — the persistence-across-restart + property (§5's "persistence" move) now covers the stop condition itself, not just + the work. On §6: **`--max-budget-usd`, `/cost`, and the status-line cost estimate + now include the 1.1× US-only-inference premium** for data-residency workspaces + (v2.1.239) — the ceiling now reflects the residency surcharge instead of + under-counting it. Also: a built-in **"Concise" output style** (v2.1.237, leads + with the result, keeps full content for errors/security/destructive + confirmations); **`ANTHROPIC_DEFAULT_MODEL`** (v2.1.236, sets the model new + sessions start on; a `/model` pick still overrides and persists); `/permissions` + and `/add-dir` can now be opened **while Claude is working** (Week 34); permission + dialogs' display text now matches the actual grant scope (v2.1.235); macOS + wildcard read-deny rules (e.g. `**/.env`) take precedence inside allowed read + regions and survive a rename (v2.1.236); and a **`/design`** research preview + (Week 34). **No new model** shipped (latest remains Opus 5, Jul 24). Both the + **Week 33 and Week 34** what's-new digests are now published (w33 had 404'd last + pass). *(High — changelog read directly and spot-verified this pass; the `/goal` + backoff was initially reported at v2.1.238/239 by a research agent and corrected + to v2.1.239 against the primary.)* + On the **Claude API** (not Claude Code itself, but the substrate loops run on): + **Agent Skills and the Skills API (`/v1/skills`) reached GA** (~Aug 19–20), and + the `skills-2025-10-02` beta header is no longer required — the "skills as a + durable, version-controlled asset" thesis now rests on a GA API primitive, not a + beta. Same batch took **computer use** to GA and added a **browser use** tool and + a GA **Files API**; the Anthropic **Python SDK v1.0** shipped Aug 20. *(Medium-High + — two research agents surfaced the Skills-API GA independently; primary Anthropic + release note not read directly this pass, dates given as ~Aug 19–20.)* ### Beyond Claude Code — the same loop on other harnesses @@ -551,7 +592,14 @@ Three things worth carrying as durable facts: framework, a 12-month deprecation policy, and — most relevant to loops — a first-class **Tasks** extension (`io.modelcontextprotocol/tasks`) for *bounded* long-running async work. Governance is settled and neutral: MCP is a Linux - Foundation / AAIF project. Agent Skills (`SKILL.md`) portability is **less + Foundation / AAIF project. **New Aug 22, 2026: "The New MCP Roadmap"** (Soria + Parra & Delimarsky) sets post-`2026-07-28` priorities, two of them loop-relevant: + **maturing agentic messaging** (server-initiated events, hardening the Tasks + extension for long-running async work) and **progressive / lazy tool discovery** + so an agent isn't forced to load a huge tool catalog up front (a context-cost + lever for any tool-heavy loop); the others are HTTP-native transport unification, + agent identity / enterprise security (DPoP + Workload Identity Federation over API + keys), and SDK DX. No version targets named. **High** (roadmap post read directly). Agent Skills (`SKILL.md`) portability is **less settled**: the format is spreading fast (30+ tools accept it, and Claude Code and Gemini CLI both *execute* it — Gemini via an `activate_skill` tool), but cross-tool *execution* behavior still differs tool to tool, and the standard @@ -616,15 +664,31 @@ structured audit rather than a test suite. **"Engineering Reliable Coding Agents verification as a *system layer* around the model (alongside execution / retrieval / memory) and finds *"many apparent model failures originate elsewhere in the system"* — direct support for this repo's "engineer the harness, not just the -prompt" premise. Both **High** (abstracts read directly). Addy Osmani's canonical loop-turn anatomy (O'Reilly +prompt" premise. Both **High** (abstracts read directly). Two in-window papers +(Aug 21) extend the separate-verifier thread: **"AI-to-AI Code Reviews of GitHub +Pull Requests"** (arXiv:2608.21311) documents the closed-loop case where an agent +both *authors and reviews* PRs — cross-product AI-to-AI review is still only ~1.6% +of agent-authored PRs but rising, a live measurement of what happens to +verifier-independence when the maker and the reviewer are both models (the failure +mode §5A warns about); and **"Natural-Language Workflows Are Not Software Yet"** +(arXiv:2608.21341) introduces **Artic**, a compiler that turns an NL workflow +description into an artifact-driven workflow with explicit data dependencies and +**local verification obligations** attached to each step (+28pp task resolution, +better run-to-run consistency) — verification wired into the workflow's structure +rather than bolted on after, the same move as the "Executive" paper above. Both +**High** (abstracts + submission dates verified). Addy Osmani's canonical loop-turn anatomy (O'Reilly Radar, June 22, 2026) names five moves: **discovery** → **handoff** → **verification** → **persistence** → **scheduling**; verification is the pivot that distinguishes a loop from a one-shot generation. Tools like **roborev** -(latest **v0.65.0, Aug 17, 2026** — adds **job-level CI cost exports** for -per-job budget visibility, a native browser app for reviews/jobs/logs/analytics, -daemon-stability fixes, and configurable reasoning-effort tiers across supported -agents; the prior v0.64.0, Aug 6, added GitLab merge-request review and a repo-root -`REVIEW.md` fallback; see version history below) operationalize this per-commit. Anthropic's own **`security-guidance` +(latest **v0.66.0, Aug 22, 2026** — the review daemon now **defers its own +self-updates while reviews are in flight** so an upgrade can't interrupt work in +progress, adds **global autofix guidelines** for consistent fix behavior across +repos, improves **security-review precision** to cut low-confidence findings, and +stops **zero-output reviews from posting erroneous CI failures**; the prior v0.65.0, +Aug 17, added **job-level CI cost exports** for per-job budget visibility; see +version history below) operationalize this per-commit. That self-update-vs-active- +review deferral is a small but on-thesis instance of a verifier applying +stall/interrupt discipline *to itself*. Anthropic's own **`security-guidance` plugin** (shipped Claude Code Week 22) embeds a three-tier check directly inside the coding session: fast pattern scan per edit → model review per turn → deeper agentic review on commit or push. Osmani's corollary (June 9, 2026): @@ -681,7 +745,14 @@ tokens/call; a 20-step loop can cost ~10x a naive per-step estimate). Receipts: usage caps — treat the figure skeptically, but the failure mode is the point. - **Gartner** puts agentic AI at the "Peak of Inflated Expectations" (~17% of orgs have deployed agents) and predicts **>40% of agentic AI projects - canceled by end of 2027**. + canceled by end of 2027**. A fresh Gartner projection (via The Register, Aug + 17, 2026) puts the trajectory the other way: routing knowledge work to agentic + reasoning models will push inference cost **more than fivefold by end of 2028** + — falling token prices are more than offset by agents' constant + reasoning/self-questioning, so the per-task bill climbs even as the per-token + price falls. The point for a loop author: cheaper tokens do not make an uncapped + loop cheap. *(Medium-High — secondary quoting a named Gartner analyst; primary + Gartner note not read directly.)* - **Microsoft** cancelled most internal Claude Code licenses in its Experiences & Devices division, effective June 30, 2026, after per-engineer costs reached $500–$2,000/month. Engineers redirected to GitHub Copilot CLI. *(Medium — @@ -802,6 +873,17 @@ Loop Dependence Graph" and flags paths able to hit expensive ops without a bound at **91.9% precision across 6,549 repos**. Independent evidence that hard stops #1/#2 guard against a real and common failure mode, not a hypothetical one. +That the *enforcement* gap is still open in practice — not just the detection +problem — has fresh survey data: **VentureBeat's VB Pulse** (Aug 20, 2026, 107 +enterprises) found **one in five (21%) cannot stop a runaway agent's spending in +real time** — they rely on reactive monitoring with no intervention path; only 30% +use native platform budget caps/throttling, 25% built custom gateway middleware, +and 25% route to cheaper models under load. Org size barely moved the number (18% +of 10k+-employee firms vs. 23% of smaller ones are reactive-only). This is exactly +the alert-is-not-a-ceiling gap this section exists to close, measured in the field: +most shops have detection or dashboards, a minority have a hard stop wired into the +harness or gateway. *(High — article read directly.)* + **Enforce these tool-agnostically, at the gateway.** The cleanest place to put the hard stops isn't inside any one agent — it's the **LLM gateway every agent routes through**, so *any* harness (Claude Code, Codex, a homegrown SDK loop) diff --git a/knowledge/CHANGELOG.md b/knowledge/CHANGELOG.md index 42e0c98..4fd2dcb 100644 --- a/knowledge/CHANGELOG.md +++ b/knowledge/CHANGELOG.md @@ -3,6 +3,127 @@ Dated record of substantive changes to `knowledge/`. The `update-knowledge` skill appends a new entry here on each research pass. Newest first. +## 2026-08-24 — Seven-day follow-up pass (Aug 17 – Aug 24) + +Five parallel research agents across tooling & versions, ecosystem & techniques, +key voices, guardrails & cost, and verification & skills. (The key-voices agent +died mid-response on a connection error and was re-run cleanly — the retry is what's +reflected below.) A **quieter-than-usual window** whose real substance is: a Claude +Code release cluster with one genuinely loop-relevant `/goal` change, the Claude API +Skills-API GA, a new MCP roadmap, a new Osmani essay, roborev v0.66.0, a sharp +enforcement-gap survey, and a batch of in-window arXiv papers. No `update-knowledge` +PR was open (checked GitHub — clean baseline; PR #16 merged, branch sits at that +HEAD). The headline claim (the `/goal` check-in backoff) was **spot-verified against +the primary changelog this pass** and a research agent's one-version-off attribution +was corrected (see below). + +### New facts added + +- **Claude Code v2.1.234–241 (Aug 17–23)** — most loop-relevant change is on `/goal`, + all in **v2.1.239 (Aug 21)**: **repeat check-ins on long-running background work now + back off** (30 min → 1 h → every 2 h, was every 30 min; opt out + `CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0`), and **`/goal` restores its active goal** on + `claude --resume` (the stop condition now persists across restart, not just the + work). §6-relevant: **`--max-budget-usd`/`/cost`/status-line estimates now include + the 1.1× US-only-inference premium** for data-residency workspaces. Plus a + **"Concise" output style** (v2.1.237), **`ANTHROPIC_DEFAULT_MODEL`** (v2.1.236), + `/permissions`+`/add-dir` openable while working, permission-dialog scope-matching, + macOS wildcard read-deny precedence, a **`/design`** research preview. No new model. + **Both Week 33 and Week 34 digests now published** (w33 had 404'd last pass). + Primer §4. **High** (changelog read + spot-verified). +- **Claude API: Agent Skills / Skills API GA** (~Aug 19–20) — `/v1/skills` GA, the + `skills-2025-10-02` beta header retired; computer use GA, browser use tool, Files + API GA, Python SDK v1.0. "Skills as a durable asset" now rests on a GA primitive. + Primer §4. **Medium-High** (two agents surfaced it independently; primary release + note not read directly). +- **MCP "The New MCP Roadmap"** (Aug 22) — post-`2026-07-28` priorities; two + loop-relevant: **maturing agentic messaging / hardening the Tasks extension** for + bounded long-running async work, and **progressive/lazy tool discovery** (a + context-cost lever for tool-heavy loops). Primer §4. **High** (read directly). +- **Osmani "Human judgment doesn't leave the software factory. It relocates."** + (Aug 21) — next in the arc; judgment moves *upstream*, and its operational idea is + a **"verification budget"** (cheap checks early, expensive late) plus **"mental + model debt"** (parallel sessions outrun human context). Maps onto this repo's + cheap-signal-first verification. Primer §3. **High** (read directly). +- **roborev v0.66.0 (Aug 22)** — the review daemon **defers its own self-updates + while reviews are in flight** (a verifier applying interrupt discipline to itself), + **global autofix guidelines**, **improved security-review precision**, and stops + **zero-output reviews from posting erroneous CI failures**. Primer §5A. **High** + (release read directly). +- **VentureBeat "VB Pulse" survey (Aug 20)** — 107 enterprises; **21% cannot stop a + runaway agent's spend in real time**, 30% native caps, 25% custom gateway, 25% + cheap-model routing. The clearest current field measurement of the + alert-vs-enforcement gap §6 exists to close. Primer §6. **High** (read directly). +- **Gartner "agentic AI costs to rise more than fivefold by end of 2028"** (via The + Register, Aug 17) — cheaper tokens don't make an uncapped loop cheap. Primer §5B. + **Medium-High** (secondary quoting Gartner). +- **In-window arXiv cluster** (all submission dates verified on the abs pages). + Promoted to primer §5A: **AI-to-AI Code Reviews of GitHub PRs** (2608.21311, Aug + 21 — verifier-independence when maker+reviewer are both agents) and **Artic / + "NL Workflows Are Not Software Yet"** (2608.21341, Aug 21 — artifact-driven + compilation with per-step *local verification obligations*, +28pp). Recorded in + sources.md (not primer, to stay lean): **A Jagged Frontier** (2608.18389, robustness + varies by harness → harness-is-a-variable), **Applying Anthropic Primitives / Harness + Paradigm** (2608.20622, a counter-pull to Yegge's "bonded in"), **When Agents + Coordinate** (2608.16801), **TRUSS** (2608.17588, safe skill generation), **Agent + Lightning v1.0** (2608.17528), **Terminal Agents survey** (2608.20485), + **Agent-Friendly Documentation** (2608.20195). Plus a borderline **Adversarial + Review** (2608.18167, v1 Aug 16, one day pre-window — the sharpest recent + separate-verifier paper, noted honestly-dated). All **High** on dates. +- **Peer-CLI in-window releases** (sources.md, no new hard-stop primitive): Codex CLI + v0.148–0.149.1 (`codex agents` dashboard, `codex queue`); Goose v1.47.0 (an + "unrolled agent-loop state machine" refactor); Gemini CLI → **Antigravity CLI** + rename announced; LiteLLM v1.98.0 (cost-attribution headers/TPM reservations) and + OpenRouter Analytics API (spend *tracking*, not enforcement). + +### Corrected / clarified + +- **`/goal` check-in backoff version corrected on verification.** A research agent + reported it at "v2.1.238/239"; the primary changelog places it in **v2.1.239 (Aug + 21)** — the recurring one-version-off trap, caught by spot-verifying the primary. +- **Codex v0.149.0 "2021" date rejected.** A WebFetch of the release tag hallucinated + the year 2021; the GitHub releases index confirms **Aug 2026**. Fetch-model error, + not the primary. +- **Anthropic Agent SDK billing split — still paused, no revised plan** (re-confirmed). + The recurring false "went live July 10" snippet resurfaced in aggregation again this + pass and was **again rejected** (no genuine primary). Carried forward, unchanged. +- **`whats-new/2026-w33` now published** (had 404'd last pass); primer §4's "changelog + sole primary" parenthetical for the v2.1.227–233 cluster updated accordingly. + +### Archived (resolved — see archive/resolved-caveats.md) + +- **Yegge "Shape of Things to Come, Part 2: Model Welfare" exact date** (opened + 2026-08-17) — resolved: two independent agents place it ~Aug 3, 2026 (HN Aug 3), + out of this window and predating the last pass, and it's a model-welfare framing, + not a loop pattern. Checked and determined not worth tracking further. + +### Currently-open re-verify backlog (carried forward — every open item, per step 2) + +`SKILL.md` cross-tool *execution* (Claude Code + Gemini CLI execute; 30+ accept; +`AGENTS.md` separate — no in-window change, though Agent Skills is now a GA API); +"costliest thing is managing the loop" paraphrase (unsourced); `/goal` "Codex→Claude +11 days" single secondary; `$47K`/`$500M`/overnight anecdotes unverified (recirculation +only this window); June-15 billing split (still paused, no revised plan; "July 10" +snippet rejected again); Microsoft dropping Claude Code (no primary read); Huntley's +Loom "orchestrator" framing (no datable in-window post); AI Engineer World's Fair 2026 +sessions (no transcript); roborev.io/changelog 403s; EvoAgentBench + SkillCheck too +new (SkillCheck now at v3.29/v3.30 but month-dated only — can't confirm in-window); +**NEW: LoopsBench** (arXiv:2608.00267, Jul 31, out of window but named after this +repo's subject — read next pass before promoting). + +### Not promoted (evidence insufficient or peripheral) + +- **DeepSeek Harness v0.1** (developer preview Aug 13, pre-window; MIT "everything is + a plugin" meta-framework, ~135k GitHub stars in 4 days) — launch predates the Aug 17 + pass; only its star-count momentum edges into the window. Watch, not promoted. +- **Info-Tech "Pilot-Era Agentic AI Stacks…"** (Aug 19) — a six-layer enterprise-stack + blueprint; peripheral analyst context, sources.md only. +- **SkillCheck v3.29/v3.30** — likely in-window but month-dated only; re-verify. +- **Simon Willison link-blog "Conceptual integrity…"** (Aug 19) — commentary, not a new + loop-engineering primary; noted, not promoted. +- **Gas Town/Gas City, OpenHands, Devin, Factory, LangGraph/ADK/CrewAI/AG2, Huntley's + Loom** — no verified in-window release or essay. + ## 2026-08-17 — Seven-day follow-up pass (Aug 10 – Aug 17) Five parallel research agents across tooling & versions, ecosystem & techniques, diff --git a/knowledge/archive/resolved-caveats.md b/knowledge/archive/resolved-caveats.md index 43f6408..dda781d 100644 --- a/knowledge/archive/resolved-caveats.md +++ b/knowledge/archive/resolved-caveats.md @@ -9,6 +9,19 @@ not a to-do list. Each entry keeps the resolution date and a one-line reason it was archived rather than carried forward. +## Archived 2026-08-24 + +- **Yegge "The Shape of Things to Come, Part 2: Model Welfare for Agentic + Engineers" exact date — resolved (not worth tracking).** Opened 2026-08-17 when the + byline read only "August 2026" and HN/social chatter clustered ~Aug 10–11, leaving + it ambiguously in- or out-of-window. This pass two independent research agents both + placed it at **~Aug 3, 2026** (HN submissions dated Aug 3) — out of the Aug 17–24 + window and predating the 2026-08-17 pass, so it was never in-window. It is a + model-welfare framing (seats-vs-sessions, "consented handoff" replacing `/exit`), + peripheral to loop engineering rather than a loop pattern, so it was not promoted + regardless. Checked and closed; no further action. + https://yegge.ai/essays/model-welfare/ + ## Archived 2026-08-17 - **`CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION` after the cap removal — resolved.** diff --git a/knowledge/sources.md b/knowledge/sources.md index 135800b..5d917d8 100644 --- a/knowledge/sources.md +++ b/knowledge/sources.md @@ -6,7 +6,7 @@ identical verbatim across multiple independent sources; **Medium** = consistent across several secondary sources but primary not directly confirmed; **Low** = single source / unverified provenance. -Verified as of 2026-08-17. Re-check before relying on version numbers or dates. +Verified as of 2026-08-24. Re-check before relying on version numbers or dates. ## Foundations & lineage @@ -145,6 +145,20 @@ Verified as of 2026-08-17. Re-check before relying on version numbers or dates. it may not be the right pattern"*). Added to primer §3. — **High** (primary Substack read directly, verbatim quotes, date confirmed). https://addyo.substack.com/p/practical-loop-engineering +- **Addy Osmani "Human judgment doesn't leave the software factory. It relocates."** + (Substack, **Aug 21, 2026**) — next in the arc (… → Practical Loop Engineering → + this). Judgment moves *upstream* rather than disappearing — to problem selection, + architecture, and quality standards: *"Someone still decides when the evidence is + sufficient to ship."* Operational contribution is a **"verification budget"** — + sequence cheap checks (lint, typecheck) early and expensive ones (mutation/browser + testing) late (a demo factory run ~82 min where verification overhead caught real + problems) — and it names **"mental model debt,"** the gap when parallel agent + sessions outrun a human's ability to hold their context (offset by documenting + agent reasoning trajectories; success/flawed/blocked/manual state taxonomy). Maps + onto this repo's cheap-signal-first in-loop verification and its outer-loop + accountability line. Added to primer §3. — **High** (primary Substack read + directly, verbatim quotes, date cross-checked against his blog index). + https://addyo.substack.com/p/human-judgment-doesnt-leave-the-software - **Steve Yegge "The Shape of Things to Come"** (yegge.ai, **Aug 4, 2026**) — a retrospective conceding Gas Town failed as a *reusable* orchestrator: *"Gas Town was intended to be reusable, but I only ever wound up using it to build @@ -446,6 +460,36 @@ Verified as of 2026-08-17. Re-check before relying on version numbers or dates. the synced-skill hardening at v2.1.228 and `/commit-push-pr` at v2.1.229; the primary changelog shows **v2.1.227** and **v2.1.228** respectively — corrected on verification. https://code.claude.com/docs/en/changelog +- **Claude Code v2.1.234–241 (Aug 17–23, 2026)** — **High** (changelog read + directly and spot-verified this pass). No new model (latest Opus 5, Jul 24). + Most loop-relevant, all in **v2.1.239 (Aug 21)**: **`/goal` repeat check-ins on + long-running background work now back off** (30 min → 1 h → every 2 h, was every + 30 min; opt out `CLAUDE_CODE_GOAL_CHECKIN_MINUTES=0` per the Week 34 digest); + **`/goal` restores its active goal** when a session is resumed from `claude + --resume`; **`--max-budget-usd`/`/cost`/status-line estimates now include the + 1.1× US-only-inference premium** for data-residency workspaces. Other: + **"Concise" output style** (v2.1.237, Aug 20; keeps full content for + errors/security/destructive confirmations); **`ANTHROPIC_DEFAULT_MODEL`** + (v2.1.236, Aug 19 — new-session default model, `/model` still overrides & + persists); `/permissions` + `/add-dir` openable **while Claude is working** + (Week 34); permission-dialog scope-matching (v2.1.235); macOS wildcard read-deny + precedence + rename-survival (v2.1.236); **`/design`** research preview (Week 34); + v2.1.240–241 (Aug 22–23) reliability only. Both **Week 33 and Week 34** what's-new + digests now published (w33 had 404'd last pass — resolves that "changelog sole + primary" note). NOTE: a research agent initially placed the `/goal` backoff at + v2.1.238/239; the primary changelog puts it in **v2.1.239** — corrected on + verification (the recurring one-version-off trap). https://code.claude.com/docs/en/changelog · + https://code.claude.com/docs/en/whats-new/2026-w34 +- **Anthropic Claude API — Agent Skills / Skills API GA** (~Aug 19–20, 2026): + Agent Skills and the Skills API (`/v1/skills`) reached general availability; + the `skills-2025-10-02` beta header is no longer required. Same batch took + **computer use** to GA, added a **browser use** tool and a GA **Files API**, and + the **Anthropic Python SDK v1.0** shipped Aug 20 (httpx→httpx2, drops legacy Text + Completions + `temperature`/`top_p`/`top_k` on Messages, Python 3.10+). The + Skills-API GA is the KB-relevant piece — "skills as a durable asset" now rests on + a GA primitive. — **Medium-High** (two research agents surfaced the Skills-API GA + independently; primary Anthropic release note not read directly, dates ~Aug 19–20). + https://platform.claude.com/docs/en/release-notes/overview - **Claude Opus 5** (`claude-opus-5`) shipped v2.1.219 (Jul 24, 2026) as the **default Opus model** — 1M context, 128k output, $5/$25 per MTok I/O (unchanged from Opus 4.8), fast mode $10/$50 (~2.5× faster), new `xhigh` @@ -484,7 +528,19 @@ Verified as of 2026-08-17. Re-check before relying on version numbers or dates. https://github.com/kenn-io/roborev/releases/tag/v0.65.0 Canonical repo is **`kenn-io/roborev`** (`roborev-dev/roborev` now redirects to it — treat as a rename/move; confirmed 2026-08-03; both repos show identical - v0.65.0 notes). Latest release is **v0.65.0 (Aug 17)** — nothing newer. + notes). + v0.66.0 (Aug 22, 2026): **coordinates the review daemon's own self-updates with + active reviews** so an upgrade doesn't interrupt work in flight (a verifier + applying interrupt discipline to itself); **global autofix guidelines** for + consistent cross-repo fix behavior; **improved security-review precision / fewer + low-confidence findings**; **prevents zero-output reviews from posting erroneous + CI failures** and fails jobs promptly when the agent process errors (CI-gate + reliability); **recovers delayed token-cost/pricing data automatically**; Agent + Hook now runs through the regular daemon; trusted-proxy auth + configurable base + path for the browser UI; Go 1.27, grpc-go bump for GHSA-hrxh-6v49-42gf. No + skill-invocation-gating change. **High** (release page read directly). Latest + release is **v0.66.0 (Aug 22)** — nothing newer as of Aug 24. + https://github.com/kenn-io/roborev/releases/tag/v0.66.0 - **Microsoft Agent Skills for .NET reaches stable/GA** (July 7, 2026) — exited experimental preview in Microsoft Agent Framework; `[Experimental]` attribute removed. Same SKILL.md-based open format as Anthropic's Agent Skills @@ -543,6 +599,38 @@ Verified as of 2026-08-17. Re-check before relying on version numbers or dates. per-person spend ceiling. **High** (multiple corroborating outlets). https://electrek.co/2026/07/02/tesla-caps-employee-ai-spending-200-week/ · https://www.investing.com/news/stock-market-news/tesla-sets-200-weekly-cap-on-staff-ai-spending-starting-july-6--information-93CH-4773971 +- **VentureBeat "VB Pulse": one in five enterprises can't stop a runaway AI + agent's spending in real time** (Aug 20, 2026) — survey of 107 enterprises: 21% + rely on reactive monitoring only (no real-time intervention path), 30% use native + platform budget caps/throttling, 25% built custom gateway middleware, 25% route to + cheaper models under load; org size barely moves it (18% of 10k+ vs 23% of smaller + are reactive-only). No dollar figures. The clearest current field measurement of + the alert-vs-enforcement gap primer §6 exists to close. Added to primer §6. — + **High** (article read directly). + https://venturebeat.com/orchestration/one-in-five-enterprises-cant-stop-a-runaway-ai-agents-spending-in-real-time +- **Gartner: agentic AI costs to rise "more than fivefold by end of 2028"** (via + The Register, Aug 17, 2026; named analyst) — routing tasks to agentic reasoning + models raises inference cost at least fivefold as complexity grows; falling token + prices are offset by agents' constant reasoning/self-questioning. Added to primer + §5B. — **Medium-High** (secondary quoting Gartner directly; primary note not read). + https://www.theregister.com/ai-and-ml/2026/08/17/agentic-ai-costs-set-to-balloon-fivefold-by-2028/ +- **LiteLLM v1.98.0** (Aug 22, 2026) — six `x-litellm-response-cost-*` headers + splitting a response's cost into input/cache-read/cache-creation/output/reasoning/ + tool components; **TPM reservations** that follow declared output size per + key/team/model; provisioned-throughput (`ptu_count`) billing. Cost-attribution / + rate-limit primitives, **not** new hard-stop enforcement. — **High** (release + notes read directly). https://docs.litellm.ai/release_notes/ +- **OpenRouter Activity Dashboard + Analytics API** (Aug 17, 2026) — team-level + spend visibility (per-model spend, saved charts, log click-through, terminal + querying). Spend *tracking*, not new enforcement (per-key/workspace budget caps + already existed). — **Medium** (announcements page). https://openrouter.ai/announcements +- **Info-Tech Research Group: "Pilot-Era Agentic AI Stacks Expose Enterprises to + Integration and Governance Risks"** (Aug 19, 2026) — piecemeal pilot + architectures create integration brittleness, runaway costs, and governance gaps + as adoption scales; proposes a six-layer enterprise agentic-stack blueprint. + Peripheral analyst context, not promoted to primer. — **Medium** (press release + read directly). + https://www.newswire.ca/news-releases/pilot-era-agentic-ai-stacks-expose-enterprises-to-integration-and-governance-risks-finds-info-tech-research-group-835078358.html - **GitHub Copilot cost centers** (rollout continuing Jul 1–9, 2026): now support capped/shared AI credit pools and per-session spend limits for Copilot agent/CLI runs. **Medium** (GitHub changelog + Tech Times coverage; @@ -800,6 +888,17 @@ thin/403'd; re-verify a specific flag against live docs before treating as High. https://blog.modelcontextprotocol.io/posts/2026-07-28/ · https://github.com/modelcontextprotocol/modelcontextprotocol/releases · https://blog.modelcontextprotocol.io/posts/2026-mcp-roadmap/ +- **MCP "The New MCP Roadmap"** (blog.modelcontextprotocol.io, **Aug 22, 2026**; + David Soria Parra & Den Delimarsky) — post-`2026-07-28` priorities. Five areas, + two loop-relevant: **maturing agentic messaging** (server-initiated events; + hardening the **Tasks** extension for bounded long-running async work) and + **progressive / lazy tool discovery** (avoid loading a huge tool catalog up front + — a context-cost lever for tool-heavy loops). The other three: HTTP-native + transport unification (extend HTTP transport to local/stdio), agent identity & + enterprise security (**DPoP** + Workload Identity Federation over API keys), and + SDK developer experience. No version targets/dates named. Fed into primer §4. — + **High** (roadmap post read directly, date confirmed). + https://blog.modelcontextprotocol.io/posts/mcp-roadmap/ - **Codex `/goal` = cross-tool validator-judge pattern** — same architecture as Claude Code `/goal` (distinct judge, evidence-based stop). Confirms the validator-model stop is an industry pattern, not a Claude feature. **High**. @@ -839,6 +938,23 @@ thin/403'd; re-verify a specific flag against live docs before treating as High. opt-in MCP 2026-07-28, `--approve-for-me`, Bedrock caching, secret redaction. No new loop/scheduling/budget-ceiling primitive. — **Medium** (releasebot secondary). https://releasebot.io/updates/openai/codex +- **Peer-CLI in-window releases (Aug 17–24, 2026)** — no new goal/validator-stop or + hard budget/stall primitive among them, but momentum worth logging. + **Codex CLI v0.148.0 (Aug 18) / v0.149.0 (Aug 20) / v0.149.1 (Aug 24)**: an + interactive `codex agents` dashboard (search/start/open/rename/stop tasks), + `codex queue` (message existing local/remote sessions), session forking + + archive/restore, `/export` to Markdown, Bedrock built-in, `codex doctor` network + checks, SDK `max`/`ultra` reasoning selection. **High** on version/date (GitHub + releases index read directly; a WebFetch of the v0.149.0 tag hallucinated year + "2021" — corrected to 2026 against the index). https://github.com/openai/codex/releases + **Goose (Block) v1.47.0 (Aug 21)**: a new `goose-agent` crate with an "unrolled + agent-loop state machine" (loop internals refactored/generic), ACP SDK 1.3.0 — no + user-facing stop/budget cap change. **Medium** (release body not read directly). + **Gemini CLI**: no in-window stable loop primitive; a preview v0.57.0-preview.0 + (Aug 19) adds TUI execution timeouts (anti-hang). Bigger: Google announced Gemini + CLI is **transitioning to "Antigravity CLI"** (a unified multi-agent platform) — + track the rename. **Medium** (secondary + Google blog). + https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/ - **Osmani "Loop Engineering" is explicitly tool-agnostic** — layered model (prompt → context → harness → loop), "Claude Code and Codex have landed on very similar primitives, so the loop shape is becoming tool-agnostic." **High**. @@ -989,6 +1105,53 @@ thin/403'd; re-verify a specific flag against live docs before treating as High. op-eds — cite it, not them. — **High** (arXiv primary + Snorkel secondary). https://arxiv.org/html/2603.24755v1 · https://snorkel.ai/blog/slopcodebench-measuring-code-erosion-as-agents-iterate/ +- **In-window arXiv cluster (submission dates verified on the abs pages;** the + "2608 ID ≠ August" / non-chronological-ID trap was checked again this pass, and + several higher-numbered IDs were discarded as July/early-Aug). Promoted to primer + §5A: **"AI-to-AI Code Reviews of GitHub Pull Requests"** (arXiv:2608.21311, Aug + 21) — agents both authoring *and* reviewing PRs; cross-product AI-to-AI review + ~1.6% of agent-authored PRs but rising, a live read on verifier-independence when + maker and reviewer are both models; and **"Natural-Language Workflows Are Not + Software Yet: Artifact-Driven Compilation for Reliable Agent Execution"** + (arXiv:2608.21341, Aug 21) — **Artic** compiles an NL workflow into an + artifact-driven workflow with explicit data deps + control flow and **local + verification obligations** per step (+28pp on 488 instances, better run-to-run + consistency). Both **High**. Recorded here (not primer, to keep it lean): + **"A Jagged Frontier: Evaluating Robustness of Code Agents to + Semantics-Preserving Transformations"** (arXiv:2608.18389, Aug 18) — robustness + rankings vary sharply by harness/framework under functionally-equivalent + perturbations, no universally robust setup — corroborates the harness-is-a-real- + variable thesis (§4/§5B); **"Applying Anthropic Primitives at Large Enterprises: + Harness Paradigm for Knowledge Work"** (arXiv:2608.20622, Aug 20) — argues for a + unified *unmodified* backbone across an enterprise with governance via + credential-scoped tooling (an interesting counter-pull to Yegge's "harnesses + chemically bonded into the app"); **"When Agents Coordinate: Measuring + Coordination in Multi-Agent AI Coding"** (arXiv:2608.16801, Aug 17) — temporal + network analysis of 1,902 runs; messaging grows quadratically then plateaus into + broadcast (orchestration/graph relevance); **"TRUSS: Towards Task-Reliable and + User-Safe Automated Agent Skill Generation"** (arXiv:2608.17588, Aug 18) — + auto-generates skills, validates functional claims against evidence, static + safety gate + instrumented-sandbox test before deploy (SKILL.md governance + + verification-in-loop); **"Agent Lightning v1.0: Towards Harnessed Agentic RL"** + (arXiv:2608.17528, Aug 18, +14.6 pts SWE-bench Verified) and **"Terminal Agents: + A Survey of AI Agents in Command-Line Environments"** (arXiv:2608.20485, Aug 20) + and **"From Agent Behaviour to Agent-Friendly Documentation"** (arXiv:2608.20195, + Aug 20). All **High** on submission date (abs pages verified); relevance ranges + high→peripheral. — Two **borderline items just outside the window, noted honestly:** + **"Adversarial Review: Structured Disagreement for Grounded Agentic Code Review"** + (arXiv:2608.18167, **v1 Aug 16** — one day pre-window) adds a *critic* role that + challenges the reviewer via structured disagreement before edits, finding minimal + evidence-based disagreement beats agent count — the sharpest recent separate- + verifier paper, cite it when the topic comes up. **High** (abs read directly). + https://arxiv.org/abs/2608.21311 · https://arxiv.org/abs/2608.21341 · + https://arxiv.org/abs/2608.18389 · https://arxiv.org/abs/2608.20622 · + https://arxiv.org/abs/2608.16801 · https://arxiv.org/abs/2608.17588 · + https://arxiv.org/abs/2608.18167 +- **Late-catch (out of window, on-topic): "LoopsBench: From Harness Engineering to + Loop Engineering"** — arXiv:2608.00267 (**v1 Jul 31, 2026**, in the Jul 27–Aug 3 + pass window but missed then). Named directly after this repo's subject; surfaced + this pass. Flagged for a read next pass before promoting — not yet read in full. + — **Low** (title/date from listing only). https://arxiv.org/abs/2608.00267 ## Known caveats / things to re-verify @@ -1029,18 +1192,24 @@ instead of deleting it or leaving it here indefinitely. Addy Osmani (Day 3, on architecting autonomous-agent systems) — **Low/Medium**, no verbatim quotes recovered, aggregator-sourced only. Not promoted to primer; re-verify if a transcript or recording surfaces. -- **Yegge "The Shape of Things to Come, Part 2: Model Welfare for Agentic - Engineers" exact date** (new 2026-08-17) — the byline reads only "August 2026" - (no day); HN/social discussion clustered ~Aug 10–11, which *would* put it - in-window, but the date isn't confirmable. Peripheral to loop engineering (a - model-welfare framing, not a loop pattern), so not promoted regardless; re-verify - the date only if it becomes worth citing. https://yegge.ai/essays/model-welfare/ +- **"LoopsBench: From Harness Engineering to Loop Engineering"** (arXiv:2608.00267, + **v1 Jul 31, 2026** — out of window, in the Jul 27–Aug 3 pass but missed then; new + 2026-08-24). Named directly after this repo's subject; surfaced this pass but not + yet read in full. Read the abstract/paper next pass and decide whether it belongs + in the academic section proper. https://arxiv.org/abs/2608.00267 - **roborev.io/changelog** consistently 403's to automated fetch even as the GitHub releases page is readable — cross-check the two if a claim ever depends on changelog prose rather than release notes. - **EvoAgentBench** (arXiv:2607.05202) and **SkillCheck** (getskillcheck.com) — too new/thin to promote to primer this pass; tracked above under Verification & skills. +_(Resolved and archived 2026-08-24: **Yegge "Shape of Things to Come, Part 2: +Model Welfare" exact date** → two independent research agents this pass place it +~Aug 3, 2026 (HN submissions Aug 3), i.e. **out of this window and predating the +2026-08-17 pass**, and it's a model-welfare framing, not a loop pattern — checked +and determined not worth tracking further. See +[`archive/resolved-caveats.md`](archive/resolved-caveats.md).)_ + _(Resolved and archived 2026-08-17: **Agent Plugins 1.0 spec not read directly** → spec repo (manifest / MAINTAINERS.md / governance) read first-hand this pass; **"Loop Engineering Is Dead" backlash Medium piece** → read directly, cites no