The market for AI agents is expanding quickly, from $7.63 billion in 2025 to a projected $182.97 billion by 2033. But the gap between planning and action remains stubborn.
An agent can reason through a task, draft a plan and still break when a menu shifts, a modal appears or a checkout form changes shape. That is the moment many enterprise teams encounter when a promising demo meets a normal web application.
Asen Lei Lei, CEO at Actionbook and a Forbes Technology Council member, has focused his recent work on that execution problem.
Reasoning is not execution
“Most failures do not happen because the model has no idea what the user wants,” Lei said. “They happen because the model cannot reliably convert that intent into the right browser action at the right moment.”
That distinction matters because the web is not a static document. It is a shifting interface of buttons, forms, overlays, permissions and page states that can change between the time an agent observes a page and the time it tries to act.
An agent that cannot reliably act can drive up costs, leave value unproven and introduce new risks.
Actionbook was built around the part of that problem Lei considers most immediate: execution reliability. He designed a deterministic action engine that translates AI intent into precise browser interactions, reducing hallucination-driven execution failures that can make agents look impressive in demos but fragile in real use.
In a live workflow, one wrong click can submit the wrong form, expose the wrong account state or send the agent into a loop that burns compute without finishing the task. That is difficult to deploy at scale.
The web needs a translation layer
The next problem is what the agent actually sees.
Screenshots and raw page text are not enough. A checkout page, travel search or business dashboard contains structure that humans infer quickly, but models often receive as noisy fragments.
A 2025 browsing-agent benchmark included 1,266 challenging problems, and human trainers solved just 29.2% within two hours.
Lei’s approach with Actionbook was to convert dynamic DOM structures into clean, machine-readable action maps before LLM interaction. The model no longer has to guess from clutter. It receives a clearer inventory of valid actions, element roles and browser targets.
“In browser automation, context is not just text,” Lei said. “It is the set of valid actions available at that moment, and the agent needs that set to be stable enough to trust.”
Efficiency decides whether agents reach production
Once the agent can act, teams face another hard limit: efficiency.
The browser-agent field is already measuring that pressure in task-level terms. A 2026 benchmark of 200 long-horizon web tasks found that the strongest frontier agents reached only 44.5% success, while trajectory efficiency remained at just 1.15%.
Those numbers point to the same production challenge Actionbook is trying to address. By reducing required browser interaction steps, lowering token consumption per task and making browser operation faster than legacy approaches, Lei’s system focuses on the cost of execution rather than asking the model to push through more retries.
Open source is becoming the adoption test
As execution layers mature, adoption will not be decided only by benchmark claims.
Developers need to inspect the system, test it on messy pages and see whether the tool behaves consistently outside a prepared demo. That is why open-source release strategy matters in this category. It turns reliability from a sales claim into something the developer community can pressure-test.
Actionbook’s launch followed that pattern, with Lei leading the public GitHub release and technical documentation.
“Open source keeps the standard honest,” Lei said. “If developers cannot reproduce the improvement on real pages, the infrastructure is not ready. Browser agents need to earn trust in public.”
The next standard is controlled autonomy
Enterprises do not only need an agent that can describe what it plans to do. They need one that can operate inside a browser while preserving sessions, limiting avoidable mistakes and keeping users in control when a task has real consequences.
The agentic AI development platform market is expected to grow from $10.75 billion in 2025 to $66.38 billion by 2031, which means more teams will be funding infrastructure for autonomous systems rather than treating agents as isolated chat features.
Actionbook fits that shift by positioning browser execution as the missing layer between LLM reasoning and real web action. Its lower browser-step count, lower token usage, open-source traction and pre-seed funding place the company in execution infrastructure, not just another wrapper around a model.
“AI agents will not become useful because they sound more confident,” Lei said. “They become useful when they can act with precision, recover from change, and give developers a browser layer they can actually depend on.”
©2026 Cox Media Group
