An Identity for Your Agent, a Skill It Learns by Watching, and a Mood Signal That Cuts Both Ways

Three stories from the last week aren't about a model getting smarter. They're about agents getting a place to stand: an identity they can be held to, a way to pick up new work without anyone writing a script, and a growing, contradictory pile of evidence about what talking to one all day actually does to a person. Here's what happened, and what's worth checking against your own stack.

1. Your Agent Can Now Sign Its Own Messages

Block, Jack Dorsey's company, released Buzz on July 21 — an open-source group chat app, built on the decentralized Nostr protocol, that puts AI agents into the same channels as human teammates instead of bolting a chatbot onto the sidebar. The pitch is a Slack-and-GitHub replacement where agents read threads, comment on pull requests, and run workflows as participants, not as a separate assistant pane. The part worth pausing on is identity: every agent gets its own cryptographic keypair, and a second signature ties each of its actions back to the human who authorized it. Buzz is explicitly model-agnostic — Claude Code, OpenAI Codex, and Block's own "goose" agents can all sit in the same channel, each independently verifiable, with a free desktop app now available for macOS, Windows, and Linux.

That two-signature pattern is the reusable idea, independent of whether Buzz itself takes off. Once an agent can post, approve, or merge something in a shared space, "who did this" stops being a log line you trust and becomes a claim you can check.

const agentAction = {
  agentId: 'pubkey:agent:buzz-goose-7f2a',
  ownerId: 'pubkey:human:jdorsey',
  action: 'approved-merge',
  agentSignature: sign(action, agentPrivateKey),
  ownerSignature: sign(action, ownerPrivateKey), // ties it back to a human
};
// Two independent signatures, not one shared session token —
// an audit trail that survives the agent's session ending.

If your product lets an agent take actions inside a shared workspace, this is worth checking against your own audit log: can you currently prove which actions were the agent's own judgment call versus a specific human's approval, or does everything just trace back to one API key?

2. Claude Learns a Job by Watching You Do It Once

On July 21, Anthropic added "Record a Skill" to the Claude Cowork desktop app: hit record, perform a task on screen while narrating what you're doing and why, stop recording, and Claude converts the demonstration — screen activity, clicks, keystrokes, and voice commentary — into a structured, reusable skill it can replay on demand. The narration is doing real work here, not just serving as a caption: saying "if the invoice total is blank, skip the row" out loud teaches an exception a silent click-track could never capture on its own. It's currently available to Pro, Max, and Team users, desktop-only for now, with an explicit warning to keep credentials and other sensitive data out of the recording.

The shift that matters isn't the screen-recording UI, it's who can now extend what an agent knows how to do. Writing a reliable prompt or script for a multi-step business workflow has always required someone comfortable writing prompts or scripts. Demonstrating a workflow you already do every week requires neither — which moves skill creation from engineers to whoever actually owns the process, at the cost of a new problem: reviewing a skill nobody wrote down in the first place.

const recordedSkill = {
  name: 'reconcile-weekly-invoices',
  source: 'screen-recording + narration',
  steps: inferredFromDemo, // clicks, keystrokes, and voice-flagged exceptions
  reviewedBy: null, // <- worth making this a required field before it ships
};

If your product is heading toward letting non-engineers teach an agent new tasks, the review step is the part to design deliberately now, before "who approved this skill" becomes a question you're asked after something breaks.

3. The Mood Data Doesn't Agree With Itself

A study of 20,847 U.S. adults published in JAMA Network Open this year, led by Harvard's Roy Perlis, found that the 10.3% of respondents who use generative AI daily had roughly 30% higher odds of reporting moderate-to-severe depressive symptoms than non-users — climbing to 50% higher among the 45-to-65 age group, with similar patterns for anxiety and irritability. It's a correlational survey, not a controlled trial, so it can't separate "AI use is contributing to distress" from "people already struggling turn to AI more" — the authors are careful about that limit themselves.

Set that next to a different, controlled body of evidence: randomized trials of chatbots used as structured, scoped mental-health support have found they measurably reduce depressive symptoms in clinical and near-clinical populations. Both findings can be true at once, and the axis that seems to separate them isn't "is it a chatbot" — it's whether the product is a bounded, protocol-driven intervention or an open-ended companion optimized for however long someone wants to keep talking to it.

const sessionDesign = {
  scope: 'bounded', // vs. 'open-ended, always available'
  onExtendedUse: (minutes) => minutes > 60 && suggestBreakOrHumanContact(),
  flagsDistressLanguage: true,
};
// The design question worth asking isn't "are we a chatbot" —
// it's which side of this object your product actually sits on.

If your product could plausibly become someone's daily emotional outlet, whether or not that was the intent, this is the week to decide on purpose whether it's built as a bounded tool or an always-open companion — the data suggests those aren't the same product wearing different skins.

What This Means for Builders

None of this week's stories is about a bigger model winning a benchmark. Buzz says accountability for agents is becoming a cryptographic property of the action, not a note in a log file. Record a Skill says the person best positioned to extend an agent's abilities is often the one who does the job, not the one who can code. And the mood data says the shape of a product — bounded and scoped versus open-ended and always-on — may matter more to the people using it than which model is behind it. Check your own stack against whichever of the three actually touches what you're shipping.

— Maya