Two things are happening in parallel this week: AI agents are getting genuinely more capable, and the people building with them are starting to notice the rough edges. The enthusiasm is real. So is the skepticism.
01
Claude Code is becoming a recruiting assistant nobody designed it to be
Cat Wu, who works at Anthropic, shared a workflow she's been running: describe the role to Claude Code, ask it to find 100 candidates with LinkedIn, Twitter, blog, podcast, and a one-line pitch for each, have it package everything into an artifact and email it over, then lock her laptop and go about her day. ---
Why it matters: Recruiting is one of those jobs where 80% of the work is information gathering that nobody wants to do. If Claude Code can handle the sourcing pass reliably, a team that used to pay a recruiter $80K/year to do top-of-funnel research is about to restructure that role. The remaining 20% (judgment calls, relationships, closing) still needs a human. But the 80% is gone.
Everyone jokes about agent overkill, but the pattern behind the joke is real
Dan Shipper, co-founder of Every, posted the week's most relatable complaint: "me: change this button color. Fable: sure I just spun up a fleet of 100 agents to get that done for you." It's a joke, but it's logging something genuine: current agent systems default to maximum force for minimum tasks. ---
Why it matters: The cost calculus on agentic AI breaks badly when a model treats every request as a reason to parallelize. If your app is invoking 100 subagents to do something that needed 1, you're burning money and adding failure points. The teams that figure out when NOT to spin up agents will build much cheaper products than the ones that just let the model decide.
The "you manage the agents, they do the work" model has a ceiling
Nan Yu argued that thinking of agent management like a real-time strategy game is a dead end. His point: AI at even modest capability levels already out-micromanages every human on the planet. Trying to supervise agents at the task level is fighting the wrong battle. ---
Why it matters: Most "human in the loop" agent designs right now assume the human is directing moment-to-moment. If that model is wrong, the interface layer most teams are building is wrong too. The more durable design is probably goal-setting and outcome review, not turn-by-turn supervision. Your product roadmap might have the wrong mental model baked in.
Amanda Askell, who works at Anthropic, made an observation that landed hard: getting a probability estimate out of a doctor is "one of life's unnecessary boss battles." She noted that even asking for a rough range gets you a shrug or liability-speak instead of a number. ---
Why it matters: This is the quiet argument for AI in medicine that nobody in healthcare marketing is making. Not "AI diagnoses cancer," but "AI will actually tell you there's a 30-40% chance this is nothing and a 15% chance it's worth a biopsy." The value isn't the diagnosis. It's the willingness to be specific.
Peter Yang, creator and product writer, had his eye firmly on the football pitch over the weekend, posting about English players needing armored escorts after a tough match result. No AI content here. Fair enough.