September 2026 Summaries

Browserbase Functions now support Secrets, a generally available feature that lets users securely store credentials such as API keys, tokens, and service logins, attach them only to selected Functions, and access them during execution through `context.secrets`. The capability allows browser automations to connect with external APIs, databases, CRMs, cloud storage, and internal services without hardcoding sensitive data in source code or passing it in every invocation. Access follows least-privilege principles because each Function receives only explicitly attached secrets, while users can view attachments without exposing values, detach access when it is no longer needed, and rotate credentials without rewriting code. Secret values are encrypted using a project-specific public encryption key before Browserbase receives them, helping keep credentials out of repositories, management interfaces, and request payloads.

Web Bot Auth is an emerging standard for allowing bots and AI agents to prove their origin through cryptographic HTTP message signatures, addressing identity verification without deciding whether a site should grant access. Built on the IETF’s HTTP Message Signatures standard, it uses asymmetric Ed25519 keys, publicly discoverable verification information, and request metadata such as expiration times and nonces to help websites confirm that a request came from a particular operator and cannot be easily replayed. Its rise follows growing automated traffic, with Cloudflare reporting in June 2026 that bots accounted for 57.4% of online requests, surpassing human traffic. While signatures can distinguish accountable operators from anonymous traffic, they do not establish an agent’s intent, safety, or reputation, leaving those judgments to websites, bot-protection services, registries, and behavioral systems. Critics warn that widely adopted registries could centralize access control despite being technically open to multiple operators, while supporters compare the approach to email authentication tools that enable reputation building without eliminating abuse. The broader challenge is creating transparent, interoperable trust systems and giving website owners effective control over which verified agents may enter and what actions they can perform.

Browserbase has joined the Cross App Access (XAA) ecosystem to help enterprises securely deploy AI agents amid rapid adoption and reported weaknesses in traditional identity controls. XAA, also called the Identity Assertion Authorization Grant, is a vendor-neutral OAuth extension incorporated into MCP that replaces persistent API keys with real-time, identity-based tokens scoped to individual tasks. Through integration with Okta, Browserbase enables agent access to its APIs under existing enterprise policies, with revocable permissions and audit logging rather than static credentials or repeated consent requests. The company positions this approach as a way to let organizations run browser-based agent automations at scale while retaining centralized visibility and security governance.

Browserbase and HUD describe infrastructure for training and evaluating web agents with reinforcement learning, emphasizing that reliable browser access, identity verification, reproducible environments, and robust grading are increasingly important as models improve. Their LiveWeb evaluation framework defines tasks through prompts, starting URLs, browser configurations, and rubrics, while QA review of execution traces helps detect reward hacking, grading errors, and benchmark-aware behavior. The discussion highlights risks from dynamic websites, CAPTCHA and bot-detection barriers, agents altering pages to mislead appearance-based graders, and models accessing public task solutions. Browserbase’s Verified mode and secure cloud browser sessions are presented as tools for enabling authorized access to sites, including authenticated workflows that protect credentials. HUD’s open-source sample tasks test capabilities such as visual counting, interpreting air-quality data, and operating web-based tools, with reported results showing uneven performance across models and sensitivity to exact-match grading, tool availability, provider failures, and possible parser bugs. The companies argue that difficult tasks with verifiable graded trajectories are especially valuable to frontier AI labs seeking training data and can become commercial assets distributed through HUD and related marketplaces.

Jev, developed by Typesafe, is presented as a “foundation model for classification” that converts unstructured application state into typed, probabilistic decisions through primitives such as Choice, Score, and Noul. Unlike traditional domain-specific classifiers or generative LLMs, Jev is intended to combine natural-language flexibility with constrained, parallelizable outputs, which Typesafe claims make it substantially faster and cheaper than LLMs for software decision tasks. Trained with reinforcement learning from calibrated decisions rather than human-preference-focused RLHF, it emphasizes confidence estimation and dependable automation over conversational output. The discussion argues that Jev is not suited to operate as a standalone agent because it lacks reasoning and generation capabilities, but can serve as a component in deterministic systems for customer-support routing, invoice processing, security triage, agent monitoring, and browser automation. In Stagehand, a browser-control framework, Jev’s Choice primitive was used to select interactable accessibility-tree nodes for actions, reducing median action latency in early tests from 1.97 to 0.46 seconds.

Browserbase argues that browser agents perform best when they write and execute code rather than rely on vision-based coordinate clicking or tightly constrained browser tools. Earlier computer-use models improved but remained vulnerable to viewport changes, hidden page elements, and long tasks that exhausted context windows, while hybrid screenshot-and-DOM approaches did not achieve sufficient production reliability. Browserbase rebuilt its open-source Stagehand SDK around a code-first, harness-agnostic Model Context Protocol interface providing browser code execution, concise DOM and accessibility snapshots, and screenshots, while retaining security controls such as domain allowlists, network protections, and policy-governed runtimes. Because models are already highly trained on Playwright syntax rather than Stagehand, the platform translates generated Playwright code into more efficient Stagehand operations at runtime, which it claims is faster and substantially more token-efficient than Playwright. The company is partnering with LangChain’s model-agnostic Deep Agents harness, which adds coding-agent techniques including planning, context compaction, and subagent delegation, to support customizable browser automation.