Workspace agents need an email surface

Workspace agents are good at summarizing docs, answering questions over a corpus, and drafting notes. They stall when the company’s knowledge still arrives as email: partner updates, customer escalations, board packets, and vendor notices. A knowledge agent without an email surface is reading yesterday’s wiki while today’s decisions land in a thread it cannot see and cannot answer.

Inbound as a knowledge source

Email is not only a send channel. It is a continuous feed of unstructured but high-signal events. A workspace agent platform should treat selected mailboxes as connectors: pull threads, extract entities, link them to projects, and propose memory writes under Ask. The agent does not need to “be the CEO.” It needs a scoped view of the streams that feed leadership decisions, with provenance attached to every claim.

Create role-owned inboxes for research, customer insights, and vendor intel. Forwarding from an executive inbox should be explicit and narrow—labels, senders, or folders—not a blanket dump of everything personal. Policy decides what the agent may read. Neuro OS keeps credentials server-side and runs extraction in a sandbox so a malicious attachment cannot become a shell session on the host.

Tag extracted facts with source message-id, date, and confidence. When two emails disagree, surface the conflict instead of averaging it away in a summary that sounds confident.

Outbound as action on insights

Once the agent has an insight—“three customers mentioned the same billing confusion”—the useful next step is often a drafted mail to support, product, or the customer success owner. That outbound is an action, not a chat reply inside the workspace UI. Default it to Ask. Show the sources that justified the draft, including links back to the originating threads. Let a human send or rewrite.

This closes a loop workspace tools often leave open: insight without ownership. The mailbox gives the insight a place to live as a conversation the counterparty already understands, and it creates an audit trail when someone later asks why a change was requested.

Do not confuse search with citizenship

A workspace can search mail archives and still leave agents without identity. Search is retrieval. An inbox is participation. Agents that only read historical exports cannot receive replies, cannot own a thread, and cannot be addressed by partners who never log into your product. Agent Inbox gives each role an address people can reply to. Forwarding into Neuro OS brings the human in when judgment is required.

Operating pattern

Start with one knowledge role and one curated forward rule. Measure precision of classification, usefulness of memory proposals, how often Ask catches a bad outbound, and how often humans still bypass the agent by emailing each other about the same topic. Expand sender allowlists carefully. Keep personal and executive private mail out of agent scope by default.

Review the knowledge role weekly the same way you would review a research hire: what did it promote into memory, what did it miss, and which forwards were noise. Retire forward rules that only create busywork.

Knowledge work compounds when the same role can observe, remember, and act under policy. Email remains how many of those observations arrive. Give the workspace agent a surface on that network, not only a better embedding index over last quarter’s exports.

Agent Inbox gives each role a mailbox people can reply to, with forwarding into Neuro OS when a human must see the thread. Outbound mail defaults to Ask. Run the role on Neuro OS. To scope the first inbox, get started.

Explore Neuro OS →