get a Claude agent to check your site every morning

Lenny's AI Builders - LAB 0018

one command sets it up. it checks your page and tells you in Slack what's broken.

i set up a Claude Managed Agent to check my Threadify landing page every morning.

the idea: make the page 1% better each day, and let that stack up.

it found a link that sends people to localhost. jeez, that's bad.

why it matters

a broken button on your page costs you people. you only see it if you click every button yourself.

i don't do that every day. nobody does.

a managed agent runs on Anthropic's computers, not yours. it can do the boring check while you build.

what you get

a small agent that checks your page and tells you what's wrong.

• a morning check. it opens your page and follows every button.
• one message. it tells you in Slack what's broken, with proof.
• a cap. you set the most it can spend each run.

the real example: my Threadify page

one command made five pieces on Anthropic's side: the agent, its sandbox, the vault and two memory stores.

the schedule is weekdays at 7:30 Perth time, with a $5 cap per run. it is still paused.

i ran it once by hand. the first run stopped: i'd saved a file pointer where its rules should have been, so it refused to act on empty rules. i fixed the rules. the second run worked.

it posted in my Slack channel and found three problems:

1. a dead end. "Open Threadify" and "Connect" both go to a page that redirects to localhost.
2. a wrong link. "See how it works" goes to pricing.
3. a mixed message. the page says two different things about X.

the run took about 90 seconds and cost 21 cents.

try it

1. run claude update, in the Claude app or in Claude Code.
2. paste the command below into Claude Code.
3. tell it the one job you want checked every morning.
4. set a spending cap for each run.
5. run it once by hand before you turn on the schedule.

copy this prompt

paste this into Claude Code. it walks you through the rest.

/claude-api managed-agents-onboard https://claude.dev/blog/building-effective-agent-automations/

you'll know it worked when

a message lands in your Slack channel that names what's broken on your page, with proof from the page.

you can see what the run cost next to your cap.

a bad result: the message is empty, or it says everything is fine on a page you know has a problem.

limits and dates

this is one setup, on my own page, on 9 October 2026.

this version finds and proposes only. in the episode, a yes in Slack does nothing yet. the fix step still needs code access and Sameera's okay.

the 21 cents is one run. your cost will depend on your page and your job.

LAB 0018 went public on YouTube on 10 October 2026 at 3:28 pm Perth time.

the rest of this LAB

• GPT-6.1 Sol Ultrafast. OpenAI says it's up to 8x faster than Sol Standard. in the API it's $12 in and $60 out per million tokens - six times Sol Standard. in Codex and ChatGPT Work it's on the Pro $500 plan. i tried it on one real task while filming: about 10 seconds before it started, then 9 minutes 57 seconds. one run, not a benchmark. my take: not 8x, and not worth six times the cost.
• Claude Dashboards and Claude Motion. ask your data a question and Claude builds a live dashboard. Motion turns reports and charts into short animations. both are in beta. Motion is on Team and Enterprise. my take: AI answers are getting more visual.
• Claude Managed Agents guide. Anthropic's guide shows an agent that reads Slack and GitHub on a schedule and posts a short brief. my take: they sell the parts so we can build the agents.
• Voyager. an open harness for video, graphics and games that can drive tools like Blender and Resolve. they say open source is coming soon. my take: the harness is the product, for now.
• Firecrawl Universal Scrape. one API for pages, PDFs, images, people and more, paid in Firecrawl credits. my take: research first, then send.
• two safety stories. Anthropic added a rule against sustained, needless cruelty toward Claude. it says normal frustration and testing are fine. three safety researchers say OpenAI fired them; OpenAI says they broke rules on handling sensitive information. we only know part of the story. my take: the Claude rule should be a recommendation, not a policy.

sources

• OpenAI Devs: GPT-6.1 Sol Ultrafast
• Tibo: day 4, instant steering
• Claude: Dashboards and Motion
• Claude Devs: automations with Claude Managed Agents
• Claude guide: building effective agent automations
• Anvisha: Voyager
• Firecrawl: Universal Scrape
• Jane Manchun Wong on Anthropic's usage policy
• Mikita Balesni: the letter
• Jasmine Wang
• OpenAI Newsroom: the reply

next step

watch LAB 0018 on YouTube - the practical starts at 11:02.

want the next one? get the next field notes. optional - these notes stay free to read.