The "one agent that does everything" PMs keep asking for

Every few weeks someone posts the same wishlist: an AI agent that monitors the feedback Slack channels, answers basic questions in feature channels, drafts the PRD from whatever inputs exist, creates the Jira tickets, posts the weekly status, and builds the exec slides. One post called it "OpenClaw for PMs." The replies split between vendor pitches and people saying it doesn't really exist yet.

Both reactions are right. Pieces of this are very real today. The whole thing, working end to end and trustworthy enough to leave alone, is not. The gap between those two is where teams get burned.

The parts that genuinely work now

• Ingesting and clustering feedback from Slack, Intercom, a CRM, tickets, transcripts — grouping noise into themes. Tedious for a human, well-suited to a model. One PM reported it saved 10–15 hours a week.
• Drafting a PRD from those inputs — given real material, a model assembles a solid first draft. (The given real material part is the difference between a useful draft and slop.)
• Generating Jira tickets from a finished spec — mechanical enough to delegate.
• Drafting status updates and summaries — squarely in the model's wheelhouse.

So the drafting and aggregating tasks are real. That alone is a serious productivity gain.

The parts that aren't ready — and shouldn't be

• Answering stakeholders unsupervised. The moment the agent says something confidently wrong, you've spent more trust than the automation saved.
• Writing to the real Jira board automatically. A spec that drifted becomes fifty wrong tickets in an instant. The agent should propose; a human should approve.
• Posting status leadership reads, unreviewed. A status update is a claim about reality. An unreviewed claim is a liability.

The pattern is consistent: the agent is excellent at producing the draft and dangerous at committing the result. Every step that writes somewhere consequential wants a gate.

Why "monitors Slack and answers questions" is the wrong thing to want

The wishlist conflates two different jobs: reducing the toil of producing artifacts and replacing the PM's judgment in front of other humans. The first is ready and valuable. The second is a trap — not because the tech is early, but because the whole point of the PM in those moments is to be accountable. An agent that autonomously answers a stakeholder has removed the one thing that made the answer trustworthy: someone who'll own it.

The honest version: an agent that does all the drafting and aggregating end to end, surfaces everything for review, and never commits to a system of record or a human audience without a person clicking approve. That's the dream with the failure modes removed.