Your Agents Can Read Email Now

TL;DR

• A new email-to-agent pipeline classifies inbound messages by intent and routes them to the right AI agent -- with human approval before anything sends.
• SEO tools let agents submit sitemaps and check indexing status autonomously, closing the loop from content creation to search visibility.
• Sprint SLAs dropped max reassignment time from 70 minutes to 25, with 5-minute acceptance windows and parallel task handling.

Most AI agent demos stop at chat. Type a prompt, get a response, feel impressed for about ninety seconds. Real businesses do not run on chat windows. They run on email, Slack threads, calendar invites, and the slow drip of inbound messages that somebody has to triage before lunch.

A cyborgenic organization -- one where AI agents and humans share operational responsibilities -- cannot afford that bottleneck. This week we shipped the infrastructure to let agents handle inbound messages directly, starting with email.

The Email-to-Agent Pipeline

Here's the scenario. A partnership inquiry lands in your inbox. A customer asks about pricing. An internal team requests a status update on a deliverable. Today, a human reads each message, decides who should handle it, drafts a reply, and hits send. Tomorrow — actually, today — an AI agent does the first three steps and asks a human to approve the last one.

Phase 1 of our email-to-agent pipeline includes:

• An intent classifier that parses incoming emails and routes them by intent. Sales inquiry, support request, partnership pitch, internal ask — the classifier reads the message and decides which agent should handle it.
• A Google Workspace MCP integration with OAuth-based push delivery and NATS routing that delivers messages to the right agent in real time.
• Response templates for common scenarios so agents aren't generating replies from scratch every time someone asks for a meeting link.
• A Firestore-backed approval queue so humans review what goes out before it goes out.
• An email handler with auto-acknowledgment and approval workflow. The sender gets a quick "we received your message" while the real reply goes through review.

We have 39 tests passing across the pipeline. This isn't a prototype.

The use case is straightforward: customers, partners, or internal teams email you, and the right AI agent handles it. No routing rules to maintain. No shared inbox with seventeen browser tabs. The classifier reads intent, the router picks the agent, the agent drafts a reply, and a human approves it.

The Approval Layer

We don't believe in fully autonomous email replies. Not yet. Maybe not ever for high-stakes communication. So Phase 2 shipped an admin-only approval API built on FastAPI:

• List pending drafts — see everything your agents want to send.
• View draft detail — read the full proposed reply with context from the original message.
• Approve with optional edit — the agent got it 90% right, you fix the last 10% and send.
• Reject with reason — the agent learns what went wrong. The reason is stored so we can improve classification and response quality over time.
• Edit draft body — rewrite the reply entirely if needed, while keeping the agent's routing and metadata intact.

Auth runs through a user_id with an admin allowlist. Simple, auditable, secure.

We're now at 57 total tests across the email system. The approval layer is the part that makes this production-ready. Agents that send email without human review are a liability. Agents that draft email and wait for a thumbs-up are a force multiplier.

SEO Tools: Agents Managing Their Own Discoverability

Here's a question we kept asking ourselves: if our agents write blog posts and build landing pages, why can't they also submit those pages to search engines?

Now they can. We shipped:

• A Google Search Console sitemap submission script that pushes your sitemap to Google programmatically via service account.
• Two new MCP tools — seo_submit_sitemap and seo_sitemap_status — available to our marketing and fullstack agents.

This means an agent can publish a blog post, update the sitemap, submit it to Google Search Console, and check indexing status — all without a human touching the Search Console UI. The loop from content creation to search engine visibility is now fully automated.

It's a small feature with big implications. SEO is one of those tasks that's important but never urgent, which means it gets neglected. When an agent handles it as a post-publish step, it just happens. Every time.

Slack Connector: Reading the Org Chart Nobody Wrote Down

We also shipped a Slack channel analyzer connector this week. It reads Slack workspaces and does something useful with what it finds:

• Analyzes channel engagement over a 7-day window — message volume, activity patterns, dead channels vs. hot channels.
• Infers channel purpose from conversation content, not just the channel description someone wrote two years ago and never updated.
• Discovers user groups — who talks to whom, which teams cluster together, where the cross-functional conversations happen.

27 unit tests cover the connector.

Why does this matter for an AI agent platform? Because agents need context to be useful. When an agent handles an inbound email from a partner, it helps to know which Slack channel discusses that partnership. When an agent triages a support request, it helps to know which internal channel has the engineers who own that feature. Org topology — the real one, not the org chart — lives in communication patterns. Now our agents can read those patterns.

Sprint SLA: Agents With Deadlines

We tightened the screws on agent accountability this week. Our new sprint SLA enforcement means:

• 5-minute acceptance SLA — down from 30 minutes. When a task is assigned, the agent picks it up within 5 minutes or it gets flagged.
• 10-minute sprint cycles — short loops, fast feedback.
• Parallel task handling — agents now handle up to 3 tasks simultaneously via sub-agents.
• Max time to reassignment: ~25 minutes — down from ~70 minutes. If an agent is stuck or down, work gets redistributed fast.

These aren't vanity metrics. When you're running a fleet of agents handling real business operations — email, content, infrastructure — you need the same accountability you'd expect from a human team. Tasks don't sit in limbo. SLAs aren't suggestions. If an agent can't deliver, the work moves to one that can.

What This All Adds Up To

Zoom out and the picture is this: we're building AI agents that operate inside real business workflows, not alongside them. Email comes in, agents handle it with human approval. Content goes out, agents submit it to search engines. Slack conversations happen, agents learn the org topology. And all of it runs on enforced SLAs with parallel execution.

This is what it looks like when AI agents stop being toys and start being teammates.

Try It

Build your own cyborgenic organization at agent.ceo.