A fifth of agentic traffic landed on ChatGPT Agent the month OpenAI retired Atlas

Share moves and volume moves point in different directions here. HUMAN reports total agentic traffic up 27% in August to a new high ^[9]. Comet's share fell nearly seven points to 40.7% ^[1], which puts July near 47.7% ^[18]. Run those together and Comet's absolute volume still grew roughly 8% ^[19]. Comet's volume grew even as its share fell; it simply took a thinner slice of a larger total.

ChatGPT Agent is the inverse. Going from 6% in July to 19.8% in August ^[2] against 27% total growth is about a 4.2x rise in volume ^[20], not the 3.3x the share change implies. HUMAN's Satori team offers one possible explanation: ChatGPT Agent absorbed traffic from Atlas, which OpenAI sunset on August 9 and folded into the ChatGPT desktop app ^[4][3].

That matters for what September's baseline looks like. Atlas closes out at 4.1% and should not appear again ^[3], so OpenAI's roughly 31% August footprint ^[6] carries forward as about 27.1% under two names instead of three ^[21]. Codex, which did not appear in the July report at all, arrived at 7.3% ^[5]. Four operators - Comet, the Claude Chrome Extension at 22% ^[7], ChatGPT Agent and Codex - cover 89.8% of what HUMAN saw ^[22]. Concentration is high. The names inside it turn over monthly, and at the tail they churn hard: Genspark 2.4% to 1.6%, Browserbase 2.6% to 2%, Browser Use Cloud 1% to 0.7% ^[8].

Read the methodology before you build policy on the mix. HUMAN identifies agents by combining behavioral signal analysis, user-agent attribution and publisher-level integration where available ^[16]. Attribution that keys on the user agent inherits whatever the client says about itself, which is the same reason an allowlist of operator names is a weaker control than a rate limit on behavior. Sector and route categorization reflects the destination endpoint, not the operator's intent ^[16], so the all-time-high 80% of activity on product and search routes ^[12] describes where agents landed, not what they were sent to do.

The route split is where the security-relevant line sits. User-account interactions rose from 6.4% to 7.1% while authentication activity fell from July's 7% peak to 5.5% ^[13][14], so account routes now draw more agent traffic than logins do ^[24]. Agents are arriving at accounts that a human already opened. Checkout and payment stayed flat at 2.1% ^[15], and e-commerce reclaimed the top destination slot at 46% of agent traffic on a 38.5% volume increase ^[10] while media fell from 43.5% to 36.2% ^[11].

HUMAN's August edition reports operator mix, destinations and route categories; it says nothing about spoofed operator identities or a malicious share of that traffic. This is capacity and policy telemetry from one vendor's platform over one month ^[16][17]. That mismatch is measurable: your July allowlist is mis-tuned by tens of points. Whether anyone attacked you through that gap is a separate question, and one this data doesn't answer.