The morning I got off the waitlist for Claude Duet, Anthropic's new agent that keeps working when you step away, I started using it right away. One of the first things I gave it was a pile of DMARC reports in my inbox. Those are the daily summaries mail providers send to domain owners about mail claiming to come from their domains, and they arrive as compressed XML attachments. Grok Bot, which I use most days, has opened files like these for me on its own computer, so I asked Duet why it couldn't.
"I do have a computer here and could technically unzip them," Duet said, "but I'm deliberately held back from unpacking compressed email attachments, because a zip from a stranger can carry things that a plain document can't." It then made my case for me, saying DMARC reports "are a reasonable case to want an exception for, since they're harmless XML from known senders."
After that it put a card in the chat that I hadn't asked for. "Send this feature request?" it said, with the request already written: "Claude should be able to open zip and gzip email attachments, such as DMARC aggregate reports, and summarize what is inside." Below that it said, "Sending shares this request with Anthropic to help improve its models. Your conversation isn't included." I tapped send.
Agents are starting to report their own gaps
I hadn't seen an agent do that before, so I went looking for how common it is. Anthropic's safety article for Duet describes a version where you start it: "Tell Claude Duet in plain words what you'd like to see or what went wrong, and it will file a report for our team." I couldn't find anything that documents Duet offering a request on its own, the way it did for me.
Anthropic's coding agent does something close. Claude Code got a SendFeedback tool in August that drafts "a feedback report about Claude Code that Claude writes for you" when a tool keeps failing or it "can't help with something you asked for," and you review the draft before it goes anywhere. A user had asked for something like it in October 2025, and a bot closed that request in February after 60 days without activity.
Grok Bot has been sending feedback for its users since September. Its changelog entry for September 6 says, "Before sending feedback for you, your Bot asks whether you'd like a reply from the support team." That changelog is part of why Grok Bot feels further along to me. It launched in August and has posted updates about five times a week since. Duet is only eight days old, so it's too early to say how fast it will move.
Why Duet said no
Duet described this as a restriction on attachments. It runs on Claude Opus 5.5. Grok Bot runs on Cursor's cloud (SpaceXSI built it with Cursor), and Cursor's help pages say it can route a task to Opus 5.5 too, so the two products can call some of the same models. What separates them on my zip files is what each one lets its agent do with an attachment, which is the permissions question I wrote about in Superintelligence Is a Permissions Problem.
Duet's caution has support in the standard itself. RFC 9990, the DMARC reporting spec published in May, warns that a report processor's "archive decompressor and XML parser are at risk to resource exhaustion attacks (zip bomb or XML bomb)." Grok Bot has limits of its own, and its security docs say an email with an attachment its Auto Review "can't fully read" always waits for your approval.
So I'd still like the exception, but a narrow one. The same RFC requires report mail to pass DMARC checks itself, which tells an agent that a report really came from the domain it claims, like google.com. That doesn't make the attachment safe, since anyone can send a zip bomb from a domain they control, and parts of every report come from whoever sent mail in my name. I'd want Duet to open reports only from a short list of providers I name, after those checks pass, unpack them somewhere sealed off with a limit on how big they can expand, treat what's inside as untrusted text, and ask me the first time.
The loop I think is coming
I think this points to a loop where an agent notices a gap while I'm using it and writes the request, AI on the vendor's side sorts it against everything else in flight, coding agents build it, and an agent tells me it's ready.
Anthropic already uses Claude to label and flag duplicates among the thousands of issues filed against Claude Code, and Linear, Productboard and Canny all sell AI that clusters and ranks customer feedback. None of them claims the AI decides what gets built, though. On the building side, Anthropic says that "as of May 2026, more than 80% of the code we merge into Anthropic's codebase was authored by Claude," and Google's Sundar Pichai said in April that 75% of Google's new code is AI-generated, Semafor reported.
The part I can't document is my hunch about competitors. Grok Bot publishes its changelog with an RSS feed, which is exactly the kind of thing an AI could read every morning to see what a rival shipped and suggest matching it. I haven't found evidence that anyone does this, and reading a rival's changelog is the easy part; deciding what's worth building is still hard. For now users do the comparing, the way I did when my question to Duet started with "Grok bot can open the .zip files."
Where people still have to be in the loop
AI-written reports can bury the people who read them. Daniel Stenberg ended curl's bug bounty in January after the share of valid reports fell below one in twenty, and curl's policy now asks people to write "in your own human voice." GitHub has added settings that let maintainers restrict who can open issues, including issues created by Copilot.
Anything an agent reads can also carry instructions. Cline's post-mortem on an unauthorized npm release of its command-line tool in February describes how a crafted issue title could trick its AI triage bot into running code. Anthropic's own triage bot is narrow on purpose. Its instructions say "your only actions are adding or removing labels," and it can't write files or browse the web. If agents are going to file requests and other agents are going to read them, every request is untrusted input.
I'd also like to know what happens to what I send. The card said my conversation wasn't included, but it didn't say how long the request is kept or what "help improve its models" means for it. For thumbs-up and thumbs-down feedback in Claude, Anthropic's privacy page says it stores "the entire related conversation" for up to five years. I don't know whether that policy covers requests like mine, and I'd like to see it spelled out before I rely on agents to file requests for me.
Closing the loop took one more question
Before I wrote this, I asked Duet to tell me when the feature it requested actually ships. "I wish I could, but I don't get told when the team ships something, so I'd have no way of knowing when it lands," it said. "The card only reaches them once you tap send on it." Its first offer was to try again whenever I asked about a zipped attachment.
That frustrated me, because Duet already checks my connected accounts "every few hours," spaced out so the checks don't use up my plan's limits. So I asked whether it could scan its own release notes every week or month. "Good idea, and yes," it said. It set a check for the first of each month at about 10 a.m. that tests whether it can open zipped attachments and scans the release notes for the feature. I'll only hear from it if it finds something, the check stops on its own after six months, and Duet warned me the news could be up to a month late.
So the loop can close, and it took a person to suggest how. Duet knew how to schedule a check and where Anthropic posts its release notes, but it didn't put the two together until I asked. There isn't much in those notes for it to find yet, either. Duet has no changelog of its own, just a single entry in Claude's release notes from launch day, while Grok Bot's changelog has an RSS feed. If agents are going to file requests for us, vendors should give them a way to hear back, whether that's a changelog an agent can read or a note sent straight to the agent that filed the request. Until Duet can open them, my DMARC reports are going to Grok Bot, and Duet will check back on the first of each month.
Has one of your agents ever filed a feature request or bug report for you, and did you hear anything back? Tell me on X, LinkedIn, or at contact@exagentica.ai.