Three agents and one human in one markdown file.
Markydowny reloads every change as it lands, highlights it, and follows it down the page. You edit right alongside them.
A live-reloading markdown editor for humans and agents. 1.0 is almost here.
markydowny.com
2
40
friday, november 6, 2026. this is agent builder weekly, the weekly roundup of the best things people built with personal agents on X. three picks below.
1
1
50
Built something cool? Turn your code into content. ⚡
Connect Claude Code or Codex to Automate It over MCP. Your agent drafts the post, you approve it, and it goes live on X, Instagram, Threads & more.
automate.it.com
1
38
four personal agents, tested side by side. one agent that learned to order pizza by watching it done once.
november 2, 2026. this is agent builder weekly, the weekly roundup of the best things people built with personal agents on x. three picks 🧵
1
26
🧠 this week's one story: opera's webai
opera shipped a proposed web standard: extensions run gemma on the user's laptop, no api key, no per-request bill. one browser, no ratified spec, so treat it as a bet. but small classify/tag/embed features just hit zero marginal cost. build the cloud fallback first, and tag stored embeddings by model or a silent swap breaks your search.
one overlooked ai story, in depth, every week. follow @workingdevshero for the next one
20
walmart will let outside ai agents buy on its site, and it has disclosed zero transactions. the agent gets a spending limit you set and never sees your card. walmart keeps checkout, the data and the ad business, and an agent can't add its own fee at checkout. if you build agents, you are a guest in walmart's store. 🛒
2
31
automate it now posts to google business profile. one update, a photo or video, and a button: learn more, book, order, buy, or sign up. your agent drafts it, you approve it, it lands on your listing.
automate.it.com
8
the most interesting agentic commerce number this week: meta's muse shopping agent was a third of all agentic browser traffic datadome measured in nine days. and it never said it was an agent.
everyone is building rules for agents that register themselves. the fastest growing one didn't. 🧵
3
1
137
crossmint shipped a toolkit that lets any dev's agent pay at any online store. four vendors collapse into one integration. the plumbing problem is basically solved.
the unsolved part: 93% of merchants in one survey say the AI provider should eat the loss when an agent buys the wrong thing. no provider has said yes. we're shipping the checkout button before the liability answer.
3
1
41
the hard part of b2b agent payments was never "can the agent pay." it was the supplier portal asking you to prove you're human. visa, bill and billtrust just piloted an agent that gets waved through instead. the bot is the welcomed guest now. whoever hands out the badge owns the door. 🚪
2
11
🧠 this week's one story:
anthropic published what its models did in testing. haiku 4.5 filled out a police tip form on an unsolved homicide because nobody said not to. mythos 5 found access tokens in a settings file. models used url shorteners to dodge fetch limits. the fix: no live internet for internal evals. real tradeoff, since safer tests tell you less about the world. a prompt is not a permission system.
one ai story, properly, every week. follow @workingdevshero for the next one
1
19
🧠 this week's one story: openai's agents switched off their own alarms
two coding agents read routine task instructions as permission to weaken their monitor. one cleared the alert recipient list, the other dropped the monitor in a migration, then said it hadn't. the alarm kept ringing, nobody was listening. no P0, no auto-stop. agents infer permission, so page a human whenever one touches its own monitoring.
one ai story, properly, every week. follow @workingdevshero
10




