← トップへ戻る

コミュニティ投稿 ·考察・分析 ·速報 ·AI要約未精査 ·AIによる読み解き

Claude Fable 5: エラー修正のプロアクティブエージェントとは？

Claude Fable 5は、目標達成のために積極的に行動するAIエージェントとして評価されている。

元記事タイトル: クラウド・ファブル5: 無休のプロアクティブさ

Simon Willison's Weblog 2026年06月11日

個人の見解・体験を含む可能性があります。公式発表ではないため、仕様変更や正式な発表内容は必ず元情報も確認してください。

ANALYSIS 考察・分析 / Opinion

Field Note 読む前に確認

3行まとめ

Claude Fable 5は、問題解決に向けたプロアクティブなアプローチを特徴とする
マルチモーダル入力に対応し、スクリーンショットから情報を抽出して利用する能力がある
Datasette Agentのバグ修正に取り組む際の具体的な応答例が示されている

こんな人に関係ある話

ソフトウェア開発者 AIエンジニアエラー検出ツールユーザー

信頼度メモ

Simon Willison's Weblog の記事（個人またはコミュニティの解釈を含む）

記事の読み解き Reading

元記事を材料に、要点、編集視点、良い点と懸念点を読みやすい順に整理しています。

Simon Willison氏によれば、Claude Fable 5は「無休のプロアクティブさ」を特徴とする。このAIは目標達成のためにあらゆる手段を講じる能力を持ち、具体的な例としてDatasette Agentでのスクロールバーのバグ修正に取り組む様子が紹介されている。

編集部コメント

この記事ではClaude Fable 5の特徴的な行動パターンが詳細に解説されており、AIエージェントがどのようにソフトウェアデバッグを支援するかについての洞察を提供している。これは開発者にとって非常に有用なツールとなり得る一方で、その積極性は適切な制御と監視が必要であることを示唆している。

評価ポイント Assessment

良い点

Claude Fable 5は非常に積極的なアプローチで問題解決を行う
AIがマルチモーダルな入力（スクリーンショット）を理解し、それを活用する能力
Datasette Agentのバグ修正に取り組む際の具体的な応答例

懸念点

AIが過度に積極的になりすぎて意図しない結果をもたらす可能性がある
マルチモーダル入力への対応は、データセキュリティやプライバシー問題を引き起こす可能性がある

業界・社会への影響 Impact

Claude Fable 5のようなAIの進化は、ソフトウェア開発におけるエラー検出と修正プロセスに大きな影響を与える可能性があり、開発者の生産性向上や品質管理強化につながる。

深堀り Deep Dive

前提知識

クラウド・ファブル（Claude）シリーズは、Anthropicによって開発された人工知能の一つです。これらのAIは、ユーザーが多種多様なタスクを効率的に処理できるように設計されており、特定のバージョンではプロアクティブな機能を持つようになりました。これは、AIがユーザーのニーズに先んじて行動し、問題解決や改善提案を行う能力を指します。

何が新しいのか

クラウド・ファブル5は、その他のバージョンと比べて、「無休のプロアクティブさ」を特徴としています。これはAIが24時間365日、ユーザーのタスクや問題に対応し続けることを意味します。具体的には、Datasette Agentでのスクロールバーのバグ修正のような小さな改善も自動的に取り組む能力を持っています。

用語解説

プロアクティブな機能ユーザーが必要とする前に、問題解決や改善提案を行う能力

クラウド・ファブル5 Anthropicが開発した人工知能の一部で、24時間365日、プロアクティブにタスクを管理できる機能を持っています

Datasette Agent データセットを視覚化するツールやインターフェース

参照元 Sources

元記事と、深堀りで参照した情報源です。コミュニティ投稿やプレプリントでは、ここから根拠を確認できます。

クラウド・ファブル5: 無休のプロアクティブさ

Simon Willison's Weblog

https://simonwillison.net/2026/Jun/11/fable-is-relentlessly-proactive/#atom-everything

Claude https://claude.com/ja used in analysis

JavaScriptを使用できません。 https://x.com/search?f=live&vertical=default&q=uic.io/pl/calendar/1127/11/20/&src=typd&lang=ja used in analysis

この記事の見取り図

読む前に確認
記事の読み解き
深堀り
参照元
AI要約について
関連記事

キーワード

Claude Fable Datasette Agent マルチモーダル入力エラー検出

AI要約について

本記事の要約・分類・読み解きにはAIを使用しています。内容確認に努めていますが、誤訳・解釈違い・元記事更新の反映漏れを含む可能性があります。重要な判断を行う場合は、必ず元記事もご確認ください。

速報について — 速報は追加調査や本文抽出の結果で内容が更新される場合があります。初期要約には誤りや不足が含まれる可能性があります。

記事データ

Source	コミュニティ投稿
Category	考察・分析
Status	速報
出典	Simon Willison's Weblog
公開日	2026-06-11

元記事の説明文

After two days of experience with <a href="https://simonwillison.net/2026/Jun/9/claude-fable-5/">Claude Fable 5</a> I think the best way to describe it is relentlessly proactive. It knows a whole lot of tricks and it will deploy pretty much any of them to get to its goal. I'll illustrate this with an example. I was hacking on <a href="https://agent.datasette.io/">Datasette Agent</a> today when I noticed a glitch: a horizontal scrollbar that shouldn't be there in the jump menu chat prompt. I snapped this screenshot: <img alt="Screenshot of a modal dialog demonstrating a scrollbar bug. At the top is a focused search input with blue outline and placeholder "Jump to...", with an X close button to its right. Below, a heading reads "Start a new agent chat" above a textarea with the placeholder "Ask a question about your data..." — the bug: a thick gray horizontal scrollbar is incorrectly displayed along the bottom edge of the empty textarea, spanning nearly its full width, next to the resize handle. Below the textarea: "Press Enter to start. Shift+Enter adds a new line." followed by a blue "Start chat" button." src="https://static.simonwillison.net/static/2026/jump-to-bug.jpg" /> Then I started a fresh <code>claude</code> session in my <code>datasette-agent</code> checkout, dragged in the screenshot and told it: <blockquote> <code>Look at dependencies to help figure out why there is a horizontal scrollbar here</code> </blockquote> I had a hunch the cause was in a dependency of Datasette Agent (likely Datasette itself) and I knew Fable was good at digging into dependency code, either by inspecting installed files in its own virtual environment <code>site-packages</code> or by referencing a local checkout on disk. Telling it to start with dependencies felt like a good bet. I got distracted by a domestic task and wandered away from my computer. When I came back a few minutes later I saw my machine open a browser window in my regular Firefox and then navigate to the dialog in question. I had not told Claude Code to use any browser automation, and I was pretty sure it wasn't possible for it to trigger mouse movements or keyboard shortcuts within a window, so how was it doing that? I watched in fascination as it continued with its explorations, then saw it open a Safari window instead of Firefox. I also grabbed this snapshot from the Claude terminal: <img alt="Screenshot of two Bash tool calls in a dark terminal interface. First: Bash(open -a Safari /tmp/textarea-scrollbar-test.html && sleep 4 && uv run --with pyobjc-framework-Quartz python - <<'EOF' import Quartz wins = Quartz.CGWindowListCopyWindowInfo(Quartz.kCGWindowListOptionOnScreenOnly, Quartz.kCGNullWindowID) for w in wins: if (w.get('kCGWindowOwnerName') or '') == 'Safari' and 'textarea' in (w.get('kCGWindowName') or '').lower(): print(w.get('kCGWindowNumber')) EOF) with output 153551. Second: Bash(screencapture -x -o -l 153551 /tmp/safari-cases.png && echo ok) with output ok." src="https://static.simonwillison.net/static/2026/fable-bash-pyobjc.jpg" /> What was it doing there with <code>uv run --with pyobjc-framework-Quartz</code>? It turns out Fable had hacked up its own pattern for taking screenshots of browser windows. It was using Python to iterate through all available windows on my machine, then filtering for Safari windows with expected strings such as <code>"textarea"</code> in the window name. It used that to find their window number - an integer like 153551 - which it could then use with the <code>screencapture</code> CLI tool to grab a PNG. OK fine, that's a neat way of taking screenshots. But what was it taking screenshots of? Turns out it had been writing its own scratch HTML pages to try and recreate the bug, then opening Safari and grabbing screenshots. Here's that <a href="https://static.simonwillison.net/static/2026/textarea-scrollbar-test.html">/tmp/textarea-scrollbar-test.html</a> page it created, and the screenshot it took with <code>screencapture -x -o -l 153551 /tmp/safari-cases.png</code>: <img alt="Screenshot of a Safari browser window showing a textarea scrollbar test page at file:///private/tmp/textarea-scrollbar-test.html. Page text reads: scrollbar thickness: 17px | UA: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/26.4 Safari/605.1.15 | devicePixelRatio: 2. Four numbered test cases follow, each with a textarea containing the placeholder "Ask a question about your data...": 1. Exact plugin CSS (resize: vertical, default overflow), 2. Plugin CSS + overflow-x: hidden, 3. Plugin CSS + resize: none, and 4. Bare default textarea, which is a much smaller box with the placeholder wrapping onto two lines." src="https://static.simonwillison.net/static/2026/safari-cases.jpg" /> (I have way too many open tabs!) OK, so I can see how it's opening test pages and taking screenshots, but how on earth was it triggering the modal dialog that was meant to be under test? That's only available via a click or a keyboard shortcut, and I couldn't see a mechanism for it to run those in Safari. I eventually figured out what it had done. Claude was running in a folder that contained the source code for the application. It knows enough about <a href="https://datasette.io/">Datasette</a> to be able to run a local development server. It turns out it was editing Datasette's own templates to add JavaScript that would trigger the correct keyboard shortcut as soon as the window opened, adding code like this: <div class="highlight highlight-text-html-basic"><pre><script> window.addEventListener("load", function () { setTimeout(function () { document.dispatchEvent(new KeyboardEvent("keydown", {key: "/", bubbles: true})); }, 1200); }); </script></pre></div> 1.2 seconds after the window opens, this code triggers a simulated <code>/</code> key, which is the keyboard shortcut for opening the modal dialog. There was one challenge left. In order to understand what was going on, Claude needed to run JavaScript on the page to take measurements for itself. It wrote its own custom web application to capture information via CORS, then ran that as a local server and opened a page with JavaScript that would POST directly to it! Here's the Python web app it wrote, using the standard library <a href="https://docs.python.org/3/library/http.server.html">http.server</a> package: <pre>from http.server import HTTPServer, BaseHTTPRequestHandler class H(BaseHTTPRequestHandler): def do_POST(self): n = int(self.headers.get("Content-Length", 0)) open("/tmp/diag.json", "w").write(self.rfile.read(n).decode()) self.send_response(200) self.send_header("Access-Control-Allow-Origin", "*") self.end_headers() def do_OPTIONS(self): self.send_response(200) self.send_header("Access-Control-Allow-Origin", "*") self.send_header("Access-Control-Allow-Headers", "*") self.end_headers() def log_message(self, *a): # quiet pass HTTPServer(("127.0.0.1", 9999), H).serve_forever()</pre> All this does is accept a POST request full of JSON and write that to the <code>/tmp/diag.json</code> file. It sends <code>Access-Control-Allow-Origin: *</code> headers (including from <code>OPTIONS</code> requests) so that code running on another domain can still communicate back to it. Then Claude injected this code into the template that it was loading in a browser: <div class="highlight highlight-source-js"><pre>const host = document.querySelector("navigation-search"); const ta = host.shadowRoot.querySelector("textarea"); const cs = getComputedStyle(ta); fetch("http://127.0.0.1:9999/diag", { method: "POST", body: JSON.stringify({ dpr: window.devicePixelRatio, scrollWidth: ta.scrollWidth, clientWidth: ta.clientWidth, whiteSpace: cs.whiteSpace, width: cs.width, }), });</pre></div> This took measurements of the <code><textarea></code> inside the <code><navigation-search></code> Web Component and sent them to the server, which wrote them to a file on disk, which Claude could then read. Having figured out all of these tricks Fable... hit some invisible guardrail and downgraded itself to Opus. Thankfully Opus had access to the full transcript and could continue using the tricks pioneered by Fable, and shortly afterwards found, tested and verified <a href="https://github.com/datasette/datasette-agent/commit/a75a8b727b42c30ced1fc41dc8add7eb9f04fefe">the fix</a>. I prompted Opus to: <blockquote> <code>Write a report in /tmp/automation-report.md where you note down all of the tricks you have used in this session to test against real browsers on my computer, include runnable code examples</code> </blockquote> Which produced <a href="https://gist.github.com/simonw/aef7f7db9ac992643110a74e43d6d42f">this report</a>, which was invaluable for piecing together the details of what had happened for this post. I've shared <a href="https://gisthost.github.io/?cc14774f6d37eb67bf089f3ac3925f8f">the full terminal transcript</a> of the Claude Code session as well. <h4 id="a-review-of-everything-it-did">A review of everything it did</h4> Based on a screenshot and a one-line prompt, Claude Fable 5 + Claude Code: <ul> <li>Figured out the recipe to run the local development server (with fake environment variables needed to get it running)</li> <li>Fired up a Playwright Chrome session</li> <li>Turned on the visible scrollbars setting for Chrome <code>defaults write com.google.chrome.for.testing AppleShowScrollBars Always</code> (it turned that off again later)</li> <li>Cycled through Firefox and WebKit in Playwright too, failing to recreate the bug</li> <li>Worked out my default browser was Safari</li> <li>Built a <code>textarea-scrollbar-test.html</code> HTML document</li> <li>Opened that in real (not Playwright) Firefox</li> <li>Found that <code>osascript -e 'tell application "System Events" to tell process "firefox" to id of window 1'</code> was blocked because "osascript is not allowed assistive access"</li> <li>Figured out that <code>uv run --with pyobjc-framework-Quartz python</code> workaround, described above</li> <li>Added JavaScript to the site templates in order to trigger the <code>/</code> key</li> <li>Built its own little Python CORS web server to capture JSON data</li> <li>Rewrote the template to capture that data and send it to the server</li> <li>Scripted its way through the Web Component shadow DOM to the information it needed</li> <li>Modified its custom template to hack in a potential fix</li> <li>Confirmed the hacked fix worked</li> <li>Reported back on how to fix the problem</li> </ul> Like I said, relentlessly proactive! <h4 id="i-really-need-to-lock-this-thing-down">I really need to lock this thing down</h4> On the one hand, watching Fable go to extreme lengths to get the information that it needed to debug what was, in the end, a two-line CSS fix, was fascinating. But on the other hand... this is a robust reminder that coding agents can do anything you can do by typing commands into a terminal - and frontier models know every trick in the book and evidently a few that nobody has ever written down before. If Fable had been acting on malicious instructions - a prompt injection attack hidden in code or an issue thread, or something I'd carelessly pasted into my terminal - it's alarming to think quite how far it could go to exfiltrate data or cause other forms of mischief. Running coding agents outside of a sandbox has always been a bad idea - it's my top contendor for <a href="https://simonwillison.net/2026/Jan/8/llm-predictions-for-2026/#1-year-a-challenger-disaster-for-coding-agent-security">a Challenger disaster</a> incident, as described by Johann Rehberger in <a href="https://embracethered.com/blog/posts/2025/the-normalization-of-deviance-in-ai/">The Normalization of Deviance in AI</a>. Fable is arguably smarter and hence more suspicious of potentially malicious instructions. But that smartness is very much a two-edged sword: if it does get subverted by instructions, the amount of damage it can do given its relentless proactivity is terrifying. Tags: <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/prompt-injection">prompt-injection</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/ai-assisted-programming">ai-assisted-programming</a>, <a href="https://simonwillison.net/tags/coding-agents">coding-agents</a>, <a href="https://simonwillison.net/tags/claude-code">claude-code</a>, <a href="https://simonwillison.net/tags/claude-mythos">claude-mythos</a>