Your AI agents spend 80% of their time waiting for you, plus three things to act on this week
Mike, from Automate & Hustle
Anthropic studied its own staff and found its Claude sessions sit idle about 80% of the time, waiting for a person to answer. That story leads, with four ways to stop being the bottleneck. Three more follow, each with steps for this week.
🧢 Automate & Hustle - This Week
⏳ Anthropic's agents wait on a human 80% of the time
🐳 DeepSeek's cheaper coding model, with free web search in Claude Code until 1 November
🖥️ Meta's 816B open model runs on two Mac Studios
🛡️ AI break-in attempts on online shops now cost about $5 each
🧩 Five habits that unblock your agents
🧰 Four tools worth a look
🎬 Six builds from my reels this week
🗞️ This Week in AI - Four Stories.
⏳ Anthropic - Its Agents Spend 80% of Their Time Waiting for a Human
On 16 October Anthropic published a study of its own staff, How employment is changing at Anthropic. Its interactive Claude sessions spend around 80% of their time idle, waiting for a person to reply. By August, the hours its agents worked in those sessions outnumbered its employees' hours six to one. One engineer put it plainly: the people are "bottlenecking the agents."
Why it matters
Reading everything its agents produce would take even non-technical staff over half their work week, by Anthropic's own estimate.
Time spent writing instructions for agents rose from 10% of interactive sessions in fall 2025 to 22% by spring 2026. Standing instructions grew from a few hundred to 8,500 in a year.
Junior technical hires fell from 27% to 9% of technical hiring, and around a third of ML managers said AI had them hiring less. Anthropic stresses this is one company's data.
How to use it
Set two review windows a day and check agent work then, not as each result lands. Skim for exceptions.
When you answer the same agent question twice, write the answer down as a standing rule in your CLAUDE.md, project file or skill.
Hand over whole jobs: a full spec plus a "done when" line, so the agent can finish without asking you.
Put a second agent in front of you as first reviewer. It checks the output against the spec before you see it.
🐳 DeepSeek - A Cheaper Coding Model, and Free Web Search in Claude Code Until 1 November
DeepSeek released V4.1-Flash-1017 on 17 October. The weights are open under the MIT licence. It finishes tasks faster with shorter answers, at an unchanged price: $0.30 per million input tokens and $1.20 per million output at peak, half that off-peak. DeepSeek also opened its own Search and Fetch APIs to developers at $0.75 per 1,000 successful calls each. Calls made through Claude Code, Codex or DeepSeek's own harness are free until 1 November, Beijing time.
Why it matters
On DeepSWE v1.1 it scores 77.0, ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0, in DeepSeek's own table.
The same table has it behind Opus 5 on Terminal-Bench 4.0 (42.3 vs 51.8) and NL2Repo (68.2 vs 75.3). It suits bulk work, not your hardest jobs.
Off-peak input is $0.15 per million tokens and cached input $0.003. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays.
How to use it
Get a key at platform.deepseek.com. The model name is
deepseek-flashand the OpenAI-style base URL ishttps://api.deepseek.com.To try it in Claude Code, set
ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic, your key asANTHROPIC_AUTH_TOKENandANTHROPIC_MODEL=deepseek-flashin a separate terminal, then runclaude.Give it repetitive jobs first: bulk edits, tests, sub-agents. Keep client data you would not hand to a third party off it.
If your agents search the web, test the Search API before 1 November while it is free through Claude Code.
🖥️ Meta - An 816B Open Model That Runs on Two Mac Studios
Meta released Muse Spark Open on 15 October. It has 816 billion parameters with 42 billion active per token, a 1-million-token context window, and takes text, images, video and audio. Meta says it is the first model in its class to run on a desktop: two M5 Ultra Mac Studios linked by Thunderbolt 5, using Apple's MLX framework.
Why it matters
On DeepSWE v1.1 it scores 60, behind the open models Kimi K3 (68) and GLM 5.3 (67). Meta's claim is "most capable US open-weight model", not best overall.
Commercial use is allowed, including fine-tuning, with a "Built with Muse Spark Open" credit. Selling it as an API above $20M revenue a year needs a separate deal with Meta.
The two-Mac setup uses the 512GB M5 Ultra, which Apple lists as coming late October with no price yet. The base M5 Ultra Mac Studio is $5,499.
How to use it
Try it hosted first. OpenRouter, Together AI, Fireworks and Groq carry it. Read each provider's data-retention terms before you send client data.
Run one real job on it and on your current model, and compare.
Only price up Macs if keeping data in-house is the point. Meta's Mac cookbook lists the steps: macOS 26.2 or later, a Thunderbolt 5 cable and
mlx-lm.
🛡️ NorForge - Criminals Point AI Agents at Online Shops for About $5 a Try
Security firm NorForge found more than 20 unsecured control panels for Cairn, an AI hacking system, being used by what it judges to be criminal groups (Inside Cairn, 19 October). One Chinese-language group ran over 1,100 attack projects across 14 panels, mostly against online shops and payment sites in the US, Japan and Australia. The group wanted footholds it could sell later. NorForge checked one claimed break-in and found the attacker's upload form live on the shop's site.
Why it matters
NorForge estimates an average attempt costs around $5 in model fees, running on DeepSeek V4.1 Flash, the cheap model from story two.
About 10% of attempts succeed, and a successful one takes a little under 23 hours on average. That is roughly ten tries, or about $50, per break-in.
The agents hunted for login SQL injection, open endpoints, file uploads and login bypasses. These are old, known holes.
How to use it
If you run a shop on Magento, WooCommerce or similar, apply the latest security patches this week and turn on automatic plugin updates.
Switch off file uploads you do not need, such as uploads on product pages, and check upload folders for files you did not put there.
Put admin logins behind two-factor, and an IP allowlist where you can.
If you run agent dashboards yourself, put them behind a login. NorForge found these panels because they were left open.
🧩 Unblock - Five Habits Worth Copying
Five habits from this week's stories.
⏱️ Review in windows. Two set times a day beats checking every result as it lands.
📌 Answer once. A question you have answered twice becomes a written rule.
📋 Spec plus "done when". Every handover says what finished looks like.
🔁 Agent checks agent. A reviewer agent screens the work before it reaches you.
💸 Route by job. A cheap model for bulk work, your best model for the hard calls.
Start with the review windows this week. They cost nothing to set up.
🧰 Hustle Toolbox
⚡ Mercury 3 - Inception's fast diffusion model for agents, voice and support bots. $0.20 in and $0.70 out per million tokens, half that at launch. For slow-feeling bots.
🔎 Exa 3.0 - Web search built for AI agents, with a new Lite tier at $4 per 1,000 requests. For builders whose agents need live web results.
🖼️ FLUX 3 Image Klein - Black Forest Labs' 5.6B open-weights model for making and editing images. For people who want image generation they control. Check the licence first.
🎬 Codex video editing - Kyle Balmer's walkthrough of cutting talking-head video by chatting to Codex. For creators sitting on raw footage.
Also
Anthropic set up a Public Benefit Fund on 18 October: 5% of its shares, subject to stockholder approval, plus 2% for every 50% rise in value above a baseline. With the co-founders' 80% pledge, it says roughly 20% of its shares now go to the public good.
In a Harvard Shorenstein Center survey of 5,000+ US adults, 29% of women had never used AI, against 22% of men.
🎬 From Mike - This Week's Builds
🔬 REA - the setup and my three X-ray prompts
✍️ /human - the structure pass skill for Claude
🚧 The three-gate check - score your app idea before you build it, plus gate four
⭐ Competitor reviews - the Claude prompt that turns 1-star reviews into your first 100 users
⚙️ My Claude Code setup - the CLAUDE.md and permissions file from the 7 moves reel
🌐 Relume Publish - the workflow and prompts behind the accountant website rebuild (sponsored)
🤝 OptiMAX - For Lead Buyers
I also run OptiMAX. We advertise under our own brands and sell each lead or live call to one buyer, never resold. If your business buys leads or calls (US accident and injury, insurance, home equity, debt; UK trades and solar), the whole spec is on one page. Read the buyer spec.
🧭 Pick Your Path - One Click
What would make AI pay off for you in the next 90 days?
🔚 Your Move
Anthropic's Claude sessions spend most of their time waiting on a person. Where do your agents wait on you the most?
Hit reply with a one-line answer. I read every one.
Mike
Automate & Hustle. The AI week in five minutes, and what to do with it.