Another browser agent enters the ring.
On August 5, Hark — the company founded by Figure AI CEO Brett Adcock — previewed Handoff, a browser use agent designed to complete tasks on websites that lack official APIs. The claim: faster execution and lower cost than existing competitors. The planned timeline: a summer release. The evidence provided so far: not much. And that’s precisely what makes this worth a careful architectural analysis rather than breathless celebration.
What We Actually Know
Let me lay out the verified facts plainly. Hark has built a computer use agent called Handoff. It operates within the browser, navigating web pages and executing multi-step tasks on behalf of users. The company asserts it outperforms GPT-5.4 and Opus 4.8 — a bold claim that immediately raises questions about methodology, benchmark selection, and evaluation criteria.
That’s essentially the full scope of confirmed information. No public benchmarks. No pricing structure. No detailed technical architecture paper. This is a preview, not a product launch, and the distinction matters enormously when we’re evaluating agent capabilities.
The Architectural Question That Matters
From a systems perspective, browser use agents face a set of challenges that are genuinely difficult to solve reliably:
- Visual grounding — correctly identifying UI elements across varied and dynamic web layouts
- Action planning — decomposing a user’s intent into a reliable sequence of clicks, scrolls, and text inputs
- Error recovery — detecting when something has gone wrong mid-task and adapting without human intervention
- Latency management — executing actions fast enough that the experience doesn’t feel worse than doing it manually
Hark’s claim that their model “predicts the next action” rather than merely predicting the next token suggests they’ve moved toward a specialized action model rather than wrapping a general-purpose language model around browser automation tooling. This is architecturally significant if true. Most current browser agents — including those built on top of large foundation models — suffer from the overhead of translating visual observations into text, reasoning in text, then converting back to actions. A model trained directly on action prediction could theoretically reduce both latency and cost.
But “could theoretically” is doing a lot of heavy lifting in that sentence.
The Benchmarking Gap
Claiming superiority over GPT-5.4 and Opus 4.8 without publishing the evaluation framework is a pattern I’ve seen repeatedly in this space. It’s not necessarily dishonest — companies often preview results before formal papers — but it makes independent assessment impossible. Which benchmarks? WebArena? Mind2Web? A proprietary task suite? The choice of evaluation set can dramatically change how agents rank relative to each other.
I’d want to see performance broken down by task complexity, domain type, and failure mode before drawing any conclusions. A browser agent that excels at simple form submissions but fails on multi-page workflows with conditional logic is a fundamentally different product than one that handles both reliably.
Cost and Speed Claims Without Numbers
The assertion that Handoff is “faster and cheaper” is meaningless without concrete figures. Faster than what baseline? Cheaper per task, per token, per API call? Are we comparing against Claude’s computer use capability, against OpenAI’s Operator, against open-source frameworks like Browser Use or LaVague?
If Hark has genuinely achieved lower inference costs through a specialized action model — one that doesn’t require the full reasoning overhead of a general-purpose LLM — that would represent a meaningful engineering achievement. Specialized models tend to be smaller, faster, and cheaper to run. But again, we need numbers to evaluate this claim seriously.
Why the API-Less Approach Matters
One detail worth examining: Handoff targets websites without official APIs. This positions it as a solution for the long tail of web services that never built programmatic interfaces — think legacy enterprise tools, government portals, small business platforms. The market for automating interactions with these systems is substantial, and existing RPA tools handle it poorly because they rely on brittle selectors and fixed workflows.
If Hark can deliver reliable task completion on arbitrary websites with reasonable accuracy, the practical value is clear regardless of whether it technically outperforms any specific foundation model on standardized benchmarks.
My Assessment
Handoff is interesting as a direction, not yet as a product. The architectural decision to build a specialized action-prediction model rather than layering automation on top of a general LLM is sound engineering if executed well. But until Hark publishes benchmarks, pricing, and reliability data, this remains a preview in the truest sense — a glimpse of intent, not a demonstration of capability. I’ll reserve judgment for when the system ships this summer and independent researchers can stress-test it against real-world tasks.
🕒 Published: