A field guide to the agent features vendors announced this quarter
June to September brought a dozen agent features from seven vendors. Sorted by what each one does, and whether it ships, is in beta, or exists only in a press release.
Press releases sort themselves by company. Engineers care what the thing does and whether it can be switched on today, so this guide is organised by function: detect, verify, respond, report, search. Each entry says whether the vendor describes it as shipping, in beta, or announced. Dates run roughly June through mid-September 2026, with two May items from the same cycle.
Detect
Plain-English custom detections (Coram, ships). Coram's homepage now leads with the ability to "build alerts in plain English" and create "custom detections across cameras." The June 17 release notes also list reduced firearm false positives and improved plate reading.
Natural-language Visual Alerts (Motorola Solutions/Avigilon, announced). At Expo Seguridad on June 2, Motorola described Avigilon Visual Alerts as using generative AI so "security personnel [can] create custom alerts for specific threats using simple natural language." The release is a trade-show preview; it does not state a ship date.
Possible fire, vehicle break-in, cash-register theft (Spot AI, announced to customers; register theft in beta). Spot AI told customers in July 2026 about three new agents: possible-fire detection, vehicle break-in, and a cash-register theft agent in beta. I could not find a public announcement page for any of the three, so treat this as a customer communication with no launch page behind it yet.
Verify
One event every 60 seconds with a written description (Ambient.ai, ships). Ambient's August 26 release says Agentic Video Walls surface "the single most relevant event once every 60 seconds, paired with an AI-generated, plain-language description of what's happening," and that the feature is now live. This is triage: the agent decides what an operator looks at, and says why. Its CTO, Vikesh Khanna, told Security Info Watch on September 16 that "Detection tells you a person is there. Reasoning tells you they have circled a loading dock three times at two in the morning."
Respond
AI deterrence on third-party cameras (Verkada, ships). Verkada's May 2026 updates extend its February talk-down feature beyond its own hardware: "Use third-party cameras to trigger a sequence of AI-generated warnings that can incorporate scene context and escalate over time if the person remains." Verkada does not use the word agent, but an escalating, context-aware voice response is what most vendors mean by one.
Agentic Response & Orchestration with Operator (Motorola Solutions, announced). From the same June 2 release: Operator "bridges the gap between detection and action, unifying data across video, access and sensor systems into a cohesive view to orchestrate and execute complex incident response." It was shown as Inform in September 2025 and renamed in March. There is no public detail on what "execute" covers without a person clicking.
Report
AI-written case narrative (Ambient.ai, ships). The same August release adds Case Management, where operators "generate a fully editable, AI-written narrative with one click, complete with auto-sorted clips and metadata." Editable is the operative word; Khanna says a person confirms the draft before it enters the record.
Deep Investigation (Coram, ships). Generally available since June. Coram's agents "analyze video and connected data across cameras, locations, and time periods to investigate what happened, identify patterns, quantify findings, and deliver a synthesized, evidence backed answer." The Series B post from June 11 describes the same agents as running in the background across cameras, access control, alarms and visitor systems.
Search
Voice mode and persistent chat (Coram, ships). Coram Assistant, the front door to Deep Investigation, gained voice input and chat history in the same release.
Conversational audit-log queries (Verkada, ships). Administrators can "filter audit logs and investigate events using conversational queries like 'show me who deleted users last month.'"
Deployment snapshots for outside agents (Rhombus, ships). The one I keep thinking about. Rhombus's July 8 post by Jeremy Gulley describes a command-line tool that exports "every location, every device, every still image" and a Claude plugin that loads it, so a general-purpose AI agent can answer "which cameras are in the east parking lot?" Rhombus is making its cameras legible to someone else's agent, and has announced no agent of its own, on the principle that "Your AI agent is only as good as the context it has."
AI Search (Milestone, announced). Milestone's March release promised natural-language search "by the end of 2026." Nothing this quarter moved that date.
What nobody announced
In the releases above, no vendor published a false-positive or miss rate for a new agent, and none described an audit log of the agent's own decisions, the record of why it talked down one person and ignored another. Nobody put a price on a feature. The nearest anyone came to saying an agent would open a door or stop a line without a person in the loop was Motorola's "execute," left undefined. Khanna's line to Security Info Watch covers the quarter: "Agentic does not mean autonomous."