Agentic Browser Automation

Enterprise agentic workflow automation attracts mega-rounds

Capital concentration in agentic automation remains intense even as deal velocity cools (velocity=-0.58), with the week of May 4 alone recording $4B deployed in a single deal. Genspark's $100M Series B extension at a $2.6B valuation and Poetic's focus on high-stakes workflows like fraud investigation and underwriting exemplify investor conviction in verticalized, enterprise-grade agents. Forward-deployed engineering teams—now backed by at least $8.5B in collective commitments from Microsoft, Amazon, OpenAI, and Anthropic—are becoming the standard enterprise GTM motion, reinforcing that the application layer is where capital is concentrating. Series B rounds account for $1.475B of the last 90 days' disclosed capital, the largest single disclosed stage bucket.

A distinct infrastructure layer for browser-operating AI agents is crystallizing, with companies solving orthogonal hard problems: Browserbase offers an open catalog of reusable SKILL.md browser automation recipes, BrowserAct handles blocked pages and authenticated sessions, and Agent Browser Shield specifically filters prompt injections and PII between agents and the live web. Coasty's 82.81% OSWorld benchmark score—achieved without APIs—demonstrates that screen-interaction computer-use agents are maturing technically. Pluno's approach of integrating directly with web-app APIs rather than UI layers signals that the infrastructure stack is already bifurcating into API-first and screen-first paradigms.

Why it matters · Infrastructure players that solve reliability and security for browser agents at scale will become critical dependencies for every application-layer automation company built on top of them.

OpenAI is no longer content as a model provider: its acquisition of Northslope signals a Palantir-style forward-deployed engineering model, embedding directly inside customer businesses. Simultaneously, ChatGPT Work launched as an explicit multi-application action-execution agent, and the GPT-5.6 model family was engineered with programmatic tool calling and multi-agent orchestration as first-class features. OpenAI also appeared as a top-5 investor in this theme with three deal-count credits, confirming it is funding the agentic application ecosystem while building competing products.

Why it matters · Startups building on OpenAI's APIs face a dual risk: the lab is simultaneously their infrastructure provider and an increasingly direct application-layer competitor.

Beijing's blocking of Meta's acquisition of Manus—despite payment already being made—is a stark demonstration that sovereign governments are now active participants in agentic AI deal outcomes. The Trump administration's case-by-case approval regime for frontier model releases adds a second regulatory layer that can gate which agent capabilities reach market. These interventions are no longer edge cases; they are structural deal risks that any cross-border agentic AI transaction must price in.

Why it matters · Cross-border M&A in agentic AI now carries material regulatory optionality costs, pushing acquirers toward domestic targets and pushing founding teams to consider domicile strategy from day one.

Hermes Agent by Nous Research has surpassed 140,000 GitHub stars by offering a built-in learning loop, persistent user modeling, 40+ tools, and full local data retention—attributes that resonate strongly with privacy-conscious enterprise developers. AutoGPT established the proof-of-concept for LLM-powered looping agents, and Goose is an early adopter of the SKILL.md standard that Browserbase is also promoting, suggesting open standards are beginning to emerge around reusable agent skill definitions.

Why it matters · Open-source frameworks with active communities create defacto standards that proprietary platforms must support or risk developer abandonment, compressing the timeline for commoditization at the agent orchestration layer.