Introducing Draft Agents

Starting today, every Draft workspace ships with agents: quiet teammates that watch the queue, link duplicates, suggest owners, and write the first draft of your release notes. They are fast, they are useful, and — this is the part we care most about — they never act without asking.

The rule we wouldn't break

Every automation feature we studied failed the same way: it did things. Issues closed themselves, labels shuffled overnight, and within a month the team had learned to distrust its own tracker. The moment a tool edits history silently, people stop believing what they read — and a tracker nobody believes is worse than no tracker at all.

So Draft Agents operate under one rule, enforced at the data layer rather than the interface: agents propose, humans decide. An agent can stage any change it likes, but nothing lands until someone accepts it. Every proposal shows its reasoning, and every acceptance is attributed to the person who made the call.

What they do on day one

• Duplicate detection. Every new issue is compared against your history — titles, bodies, and linked code paths. Likely matches surface as a proposal with the reason the two were paired.
• Suggested triage. New issues arrive with a proposed owner and severity, learned from how your team has routed similar work before.
• Release notes. When you cut a release, an agent assembles a first draft of the notes from the merged work — grouped, deduplicated, and written in plain language.

The review queue

Proposals would be exhausting if they scattered across the app, so they don't. Everything an agent wants to do lands in a single review queue. J and K move through it, enter accepts, X dismisses. A morning's triage for a hundred-person workspace takes one coffee.

Trust in a tracker is compound interest. Agents that ask permission earn a little every day; agents that act silently spend it all at once.

What we're not doing

We are not training on your workspace by default, we are not letting agents touch anything security-sensitive, and we are not adding an agent that argues in your comment threads. Each team chooses which agents run and who reviews their proposals — including the choice to run none at all.

Agents are live today for every Pro and Business workspace. Open the command menu, type agents, and flip on the ones you want. Then tell us what they got wrong — the review queue has a feedback action wired straight to the team that builds them.