Sep 9, 2026 | RUMCG Meeting

Agenda

• Keeping an Eye On:

• Declarative Performance Observer
• NavigationTiming landing attribute

• Barry's Bytes
• ABC revisit - Agentic Browsers

Admin

• Next meeting: October 14, 2026

Key Takeaways 

• Agentic browsers complicate detection: They mimic humans and popular user agents, making traditional methods ineffective and data on their traffic is lacking.
• Detection standards lag behind: Few agentic browsers self-identify; no standard API exists, hindering transparency and measurement efforts.
• RUM data impact uncertain: Ad/script blocking by agents skews metrics; traditional performance measures may lose relevance for non-human viewing.
• Business impact rising: Retailers see agent use growing; identifying agent sessions is vital for accurate SEO and performance analysis.
• Performance measurement needs evolving: Shift focus from browser speed to overall user experience across devices including AI agents.
• Community split on advocacy: Some doubt voluntary self-ID due to privacy/incentive issues; others urge pushing for standards to improve visibility.

Minutes

Agentic Browsers and Detection Challenges

The group explored the emerging category of agentic browsers, highlighting the complexity of detecting their presence and understanding their behavior on websites.

• Agentic browsers blur human and automated interactions, complicating detection and measurement (00:05)

• These browsers range from AI-enhanced traditional browsers to fully autonomous agents controlling navigation and actions.
• Many agentic browsers mask themselves by mimicking popular user agents like Chrome, making standard detection methods ineffective.
• Detection attempts focus on behavioral patterns, header inconsistencies, TLS fingerprinting, or referer changes, but no definitive public data exists on their traffic volume.
• Nic Jansma emphasized the difficulty in distinguishing agentic browsers from genuine users, especially when the agents behave like humans.

• Current detection methods and standards fall short, raising questions about the need for self-identification (00:15)

• Most agentic browsers do not modify the user-agent string to signal their AI nature except for a few like Google Agent and Manus AI.
• The WebAuth specification offers a public-key based verification system primarily designed for crawlers rather than agentic browsers.
• The lack of a JavaScript API or standardized client hints for agent identification complicates measurement and transparency.
• Discussion participants debated the potential and feasibility of advocating for agentic browsers to self-identify through user-agent tokens or headers.

• The impact on Real User Measurement (RUM) data is uncertain but likely significant due to ad-blocking and script blocking by some agentic browsers (00:19)

• Agentic browsers often block ads and tracking scripts, resulting in underreported traffic and skewed analytics.
• This raises questions about whether performance metrics like page load time or layout shifts remain relevant if no human is actively viewing the page.
• The group debated if agents would care about performance or user experience metrics traditionally tracked, or if new metrics are needed.

• Retail and SEO perspectives highlight the urgency to understand agentic browser behavior and its effect on business outcomes (00:25)

• Magnus Dahl pointed out that retailers see a shift where users increasingly rely on AI agents instead of traditional search, requiring sites to be agent-friendly.
• Ryan Townsend stressed that co-browsing scenarios where agents actively interact with sites should encourage agent self-identification to maintain accurate RUM data.
• Barry Pollard noted that some agent activity is more of a backend or SEO concern than strictly a RUM issue, emphasizing the evolving ways agents access content, including APIs replacing UI scraping.

Strategic Implications for Web Performance and User Experience

The meeting framed agentic browsers as a major shift in how users interact with the web, calling for a broader rethinking of performance measurement focused on the speed of user experience rather than just browser-based web performance.

• Sergey Chernyshev proposed shifting focus from web performance to measuring overall speed of user experience across all platforms (00:45)

• He traced the history from traditional web browsing to mobile devices and now to AI-driven interactions, emphasizing the evolving software layer on users’ devices.
• He argued that RUM should abstract beyond browsers to measure user experience regardless of delivery method, including native apps and agentic browsers.
• Sergey suggested that user experience metrics should remain central, but their collection methods and scope may need to adapt to new technology paradigms.

• The group agreed measuring agentic browser impact on key business KPIs remains important but challenging (01:00)

• Ivailo Hristov emphasized the need to reliably identify agentic browsing sessions first to correlate performance with task success rates like checkouts.
• Measuring speed and stability remains relevant since agents require stable environments to complete automated tasks, such as filling forms.
• Sergey summarized that speed is a proxy for good user experience and ultimately ties back to business outcomes, regardless of the technology delivering it.

• The conversation underscored the fluid and rapidly evolving nature of agentic browsers, complicating long-term measurement strategies (00:40)

• Barry Pollard highlighted uncertainty about the longevity and standardization of agentic browser identifiers given the fast pace of technology change.
• Participants expressed skepticism about widespread adoption of self-identification due to privacy concerns, technical challenges, and incentives misalignment.
• Matthias Krzeszowiak warned that any signals or identifiers introduced might be spoofed, reducing their usefulness and complicating data trustworthiness.

• Declarative performance measurement and new APIs were suggested as potential technical avenues to adapt to agentic browsing (00:32)

• Cliff Crocker proposed exploring Declarative Performance Observer to collect telemetry without relying on JavaScript execution, potentially useful for agentic contexts.
• Nic Jansma agreed this could be a promising area to brainstorm further, indicating openness to evolving measurement techniques.

Community Perspectives on Advocacy and Next Steps

The group debated whether the RUM community should actively push for standards around agentic browser identification and measurement or take a wait-and-see approach.

• Concerns about the futility of advocacy given adversarial incentives and evolving techniques (00:24)

• Sergey Chernyshev compared the situation to bot detection arms races, doubting that agents would voluntarily self-identify unless incentivized.
• Matthias expressed pessimism that only “good actors” would comply, while “bad actors” would evade detection, limiting impact.
• Karlijn Lowik argued that silence would be worse, advocating the community should still speak up to create some norms rather than give up.

• Some see clear benefits in standardizing agent identification for visibility and optimization (00:25)

• Magnus Dahl advocated for simple detection methods like user-agent headers to better understand agent behavior and optimize websites accordingly.
• Ryan Townsend emphasized identification especially in co-browsing scenarios to maintain accurate user and performance metrics.
• Nic Jansma suggested incentives such as browsers wanting to report their agentic usage publicly or websites serving optimized content to agents.

• Diverse views on how to categorize and respond to different agentic interactions (00:36)

• Barry Pollard outlined three agent categories: co-browsing assistive agents, crawling agents scraping data, and fully autonomous agents acting in the background.
• He noted the complexity in reporting and measurement when agent activity mixes with human browsing.
• The group acknowledged the need to differentiate these categories when considering measurement and standards.

• Ideas for behavioral or rate-limit triggers to detect non-human patterns were proposed (00:44)

• Shubham Gupta suggested a “step counter” style approach where websites monitor unusual browsing behavior volumes and prompt for bot detection or rate limits.
• This could provide a reactive mechanism for flagging agentic or automated traffic without relying solely on explicit self-identification.

Technical and Measurement Considerations

The discussion touched on technical nuances in how agentic browsers interact with websites and the implications for measurement accuracy and instrumentation.

• Agentic browsers use diverse methods to parse and interact with pages, affecting detection feasibility (00:11)

• Some agents crawl the accessibility tree or DOM to extract content precisely, while others rely on screenshots and image analysis.
• Agents may also use simple HTTP fetches or headless browsers, each leaving different fingerprints.
• This diversity complicates defining a one-size-fits-all detection or measurement approach.

• Agentic browsers often disable or mask typical bot detection signals (00:16)

• They frequently hide WebDriver flags, modify JavaScript APIs, or suppress client hint headers to evade detection.
• These factors reduce the effectiveness of traditional bot detection methods.

• Network and TLS fingerprinting might offer additional clues but are limited in reliability (00:16)

• Agents running on residential IPs with standard Chrome TLS stacks are much harder to distinguish than cloud-based headless browsers.
• The absence of standardized signaling makes behavioral analysis more important but also more complex.

• The lack of public data on agentic browser traffic limits ability to assess impact (00:18)

• Nic Jansma noted no available public statistics on the volume or share of traffic generated by agentic browsers.
• The discontinuation of some agentic browsers may suggest limited adoption but also consolidation into other products.

• Potential for websites to adapt content delivery based on agent signals to optimize performance and resource use (00:42)

• If agents self-identified, sites could serve simplified or markdown-style data to reduce CPU and bandwidth consumption.
• This would create incentives for agents to identify themselves to improve efficiency and user experience.

Meeting Logistics and Future Planning

The group concluded with logistical notes and plans for upcoming meetings.

• Next RUM Community Group meeting scheduled for October 14, ahead of W3C TPAC (00:04)

• This meeting provides an opportunity to discuss and coordinate topics relevant to agentic browsers and interoperability.
• Members were encouraged to propose agenda items for the next session.

• Recording and participation protocols confirmed (00:02)

• The meeting was recorded with participant consent, allowing review and documentation of discussions.
• New participants were welcomed and invited to contribute to ongoing brainstorming.

• Thanks and closing remarks emphasized the value of ongoing dialogue (01:02)

• Nic Jansma thanked participants for their input and encouraged continued engagement on this rapidly evolving topic.
• The session ended with agreement to revisit these discussions as technology and community needs develop.

Action items

Nic Jansma

• Share meeting recording and presentation deck related to Agentic browsers with the group (03:43)
• Prepare agenda for the October 14 meeting, focusing on Agentic browsers and potential interop topics (04:55)
• Open a ticket to brainstorm use cases for Declarative Performance Observer related to agentic browsers (33:35)

Sergey Chernyshev

• Prepare and deliver a follow-up presentation expanding on the concept of speed of user experience beyond browsers, including native apps and agentic agents (45:15)

Group

• Continue collecting community input and experiences about detection and impact of agentic browsers on Real User Monitoring data (20:35)
• Explore possible advocacy or standards proposals for agentic browsers self-identification mechanisms such as user agent tokens, client hints, or JavaScript APIs (20:35)

Shubham Gupta

• Consider ideas for behavioral detection or rate-limiting approaches based on non-human request patterns to identify automated browsing agents (44:50)

Karlijn Lowik

• Provide further retail use cases and feedback on preferred detection methods for agentic browsers (25:30)