Summary: Ocean, an agentic email security platform, runs over one billion monthly investigations of incoming emails, using Tavily as the web research layer that verifies senders in real time. Per Tavily’s October 14, 2026 case study, integration took weeks and investigation time dropped from minutes to seconds. For builders, it’s a useful worked example of buying versus building agent infrastructure.
The problem is context, not malware signatures
The hardest phishing attacks Ocean’s agents target don’t look technical. Business email compromise, vendor impersonation, and AI-written spear phishing often arrive from senders with clean reputations, so the deciding question is contextual: is this sender who they claim to be, and does the business story in the email hold up?
Answering that requires live information about people, companies, domains, and business relationships — data that only exists on the open web. According to Bar Tzemach, Security Researcher at Ocean, the team needed agents to build a sender profile and check claims inside the email in real time, at a scale of hundreds of thousands of mailboxes.
What they bought instead of built
The interesting decision here is subtraction. Ocean did not build its own web crawling and research layer; it integrated Tavily, which Tzemach describes as built with agents in mind and easy to plug into their agentic ecosystem. The move from first test to full production took a matter of weeks, per the case study, and the team kept its engineering focus on detection logic and agentic intelligence — the part that differentiates the product.
This is the same sizing question most agent teams eventually face, whether it’s search, embeddings, or reranking: spend your scarce engineering time on the layer that’s actually yours. I wrote about a similar framing for retrieval infrastructure in sizing your embedding stack by request shape, not request count — the shared principle is that shared commodity components are usually worth buying, while your detection or business logic is worth owning.
What one investigation actually looks like
The concrete example from the case study is worth unpacking. An employee was in an existing conversation with a legitimate vendor. A second email arrived from a slightly altered version of that vendor’s domain. Ocean’s agent used Tavily to research the address — who the entity claimed to be, where it operated, how it presented itself online — flagged the mismatch as vendor impersonation, and presented its reasoning to the customer’s security team. The threat was blocked and removed within seconds, with no human intervention.
Note the shape of that flow: the web search result isn’t the verdict. It’s evidence the agent pairs with proprietary threat intelligence and behavioral signals before deciding. The verdicts are grounded in verifiable sources, which Ocean says matters for building trust with customer SOC teams.
The numbers, and the constraint to watch
Per the case study, Ocean’s agents flagged roughly 54,000 malicious emails and caught about 900 lookalike or impersonation domains in September 2026 alone. Customer count grew 1.5x in four months while investigation volume nearly doubled — with no added web-research workload on humans in the loop, which the company attributes to the relevance of Tavily’s results.
That last claim is the one I’d treat as vendor-reported rather than independently verified; the supplied material is a Tavily customer story, so metrics are framed favorably by definition. Still, the operational takeaway holds: when your agents’ web layer returns results precise enough that nobody re-checks them manually, scale stops being a staffing problem.
Where this goes next
Ocean is now exploring Tavily beyond detection — meeting prep for Sales, competitor analysis for Product, marketing materials. Tzemach calls the search layer “our agents’ eyes on the open web.”
The practical step for your own roadmap: audit where your agents make claims about the outside world, and ask whether that knowledge is fetched live or assumed. If a single vendor query can impersonate a domain your customer already trusts, static training data isn’t enough. Start with the one decision where stale context costs the most, and wire search in there first.