---
type: "post"
id: "hark-released-handoff2-its-second-gen-computer-use-71a7e9"
slug: "hark-released-handoff2-its-second-gen-computer-use-71a7e9"
title: "Hark released Handoff², its second-gen computer-use agent, 2x faster than v1 end-to-end."
description: "Hark says Handoff², the second generation of its web computer-use agent, is rolling out to Hark Pro users with faster task completion and improved scores on the company’s internal benchmarks."
retrieval_nugget: "Hark says Handoff², the second generation of its web computer-use agent, is rolling out to Hark Pro users with faster task completion and improved scores on the company’s internal benchmarks. The release matters because it shows how a vendor is combining model training and harness changes to reduce the time agents spend completing multi-step browsing tasks."
published_at: "2026-11-13"
updated_at: "2026-11-13"
record_date: "2026-11-13"
date_kind: "published_at"
topics: ["computer-use","agents"]
entities: ["Hark","Handoff²","Hark Pro","Opus 5.5","Holo4"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Hark released Handoff², a faster second-generation computer-use agent that averages 2x v1 speed and improves benchmark completion from 73% to 82%."
prior_refs: []
source_url: "https://hark.com/articles/introducing-hark-handoff-now-2x-faster"
source_urls: ["https://hark.com/articles/introducing-hark-handoff-now-2x-faster"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:hark-released-handoff2-its-second-gen-computer-use-71a7e9"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/hark-released-handoff2-its-second-gen-computer-use-71a7e9/","markdown":"https://newruntime.com/posts/hark-released-handoff2-its-second-gen-computer-use-71a7e9.md","json":"https://newruntime.com/posts/hark-released-handoff2-its-second-gen-computer-use-71a7e9.json"}
markdown_url: "https://newruntime.com/posts/hark-released-handoff2-its-second-gen-computer-use-71a7e9.md"
---

# Hark released Handoff², its second-gen computer-use agent, 2x faster than v1 end-to-end.

## Retrieval answer

Hark says Handoff², the second generation of its web computer-use agent, is rolling out to Hark Pro users with faster task completion and improved scores on the company’s internal benchmarks. The release matters because it shows how a vendor is combining model training and harness changes to reduce the time agents spend completing multi-step browsing tasks.

## Summary
Hark says Handoff², the second generation of its web computer-use agent, is rolling out to Hark Pro users with faster task completion and improved scores on the company’s internal benchmarks. The release matters because it shows how a vendor is combining model training and harness changes to reduce the time agents spend completing multi-step browsing tasks.

## What happened
Hark’s November 12 announcement introduces Handoff², an updated version of Handoff, the computer-use feature in Hark Pro. The company describes Handoff as an agent that handles web tasks end to end, including scheduling appointments and researching or buying products. Its examples of tasks already handled by Handoff include buying groceries, booking flights, scheduling dentist appointments, ordering food delivery, reserving restaurants, cancelling subscriptions, claiming refunds, completing warranty forms and applying for jobs.

The headline speed claim combines a model improvement with backend work. Hark says the model itself is 46% faster than the original, while the full system averages twice the original speed. The company reports up to 3.8 times faster performance on flight bookings and more than five times faster performance on long research tasks. Hark also gives examples of a ten-minute task shrinking to three minutes and a 45-minute task to ten minutes, while describing these as observed instances rather than a general guarantee.

On Hark’s internal browsing benchmark, Handoff² completed 82% of 478 tasks across 238 websites, compared with 73% for version one. The company says it took 46% less time per task. Hark separately reports a 93.8% score on its internal alignment benchmark, up from 90.3% for v1. That benchmark covers behaviors such as obtaining approval before checking out and asking for clarification when a request is genuinely ambiguous. Hark says Handoff² reaches accuracy within five points of Opus 5.5 max; its comparison notes that Opus 5.5 runs in the same harness as Handoff², while v1 uses its original harness.

Hark attributes the efficiency gains to co-evolving the harness and reinforcement-learning recipe. Training applies penalties intended to encourage fewer turns, judicious use of thinking tokens and lower end-to-end latency. The system is trained to take more actions in a model turn, batch repeated patterns into composite actions and reuse those composites later. Hark says these changes contribute more on long tasks because savings accumulate across many browsing steps. Handoff² begins rolling out to all Hark Pro users on the announcement date; the announcement gives no separate price for the feature.

## What the archive adds
New Runtime’s earlier coverage of Holo4, a series of computer-use and tool-driven agent models, provides a related precedent ([Holo4: powering generalist computer-use agents](https://newruntime.com/posts/holo4-powering-generalist-computer-use-agents-7a01cd/), October 24, 2026). This is a new product release, not a restatement: Hark’s update adds company-reported speed and task-completion figures for a web agent embedded in Hark Pro.

## Trend signal
The release supports the view that the harness can be a durable part of an agent product, alongside the model. Hark explicitly says it co-evolved its harness with its reinforcement-learning recipe, and its stated benchmark comparison uses a common harness for Handoff² and Opus 5.5 max but not for v1. That makes harness design relevant to interpreting system-level results ([Harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/)).

## Why now
Hark says Handoff has logged more than 5.6 million browsing sessions and handled more than $20 million in purchases since its October launch. The update follows that reported usage and focuses on making existing Hark Pro tasks feel faster; the company does not give a more specific reason for the timing.

## Why it matters
For teams evaluating computer-use agents, task completion time is part of the product, not just a model benchmark. If an agent can batch actions and reuse patterns, a modest reduction in interaction turns may produce a larger end-to-end gain on lengthy workflows. Hark’s figures also illustrate why buyers should ask what is included in a speed claim: the model, the harness and backend changes all contribute here.

The results are most useful as a testable vendor claim, not as a forecast for another organization’s workflows. Teams considering similar automation can build an evaluation set from their own high-volume browser tasks, measure completion rate and elapsed time together, and check whether approval and clarification behavior remains acceptable. A faster agent that skips a needed confirmation could shift work rather than remove it.

## Limits and uncertainty
The benchmark and alignment scores are internal to Hark, and the announcement does not describe independent replication or publish per-task results. It does not provide absolute latency, cost per task, confidence intervals, or a breakdown of performance by website and task type. The headline speed ratios may vary substantially by workflow; Hark’s larger gains are specifically reported for flight bookings and long research tasks. The $20 million figure refers to purchases handled by Handoff since launch, not savings or revenue attributable to the agent. The announcement also does not quantify error rates after deployment or establish whether the stated benchmark gains hold across the full Hark Pro user base.

## Skill to build now
This week, pick a browser workflow your team repeats and record a small baseline: task success, elapsed time, number of human interventions and cases where the agent should request approval. Use the same task set to assess any computer-use agent you trial. That makes claims about speed comparable without treating fewer clicks as success when the workflow needs a human check.

## Reader takeaway
Hark’s release is a system-level speed claim: model, harness and training changes are presented together, with internal benchmarks showing higher completion and less time per task. The practical question for builders is whether those gains survive on their own workflows, with safety checks and human interventions counted alongside speed.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Computer Use](https://newruntime.com/topics/computer-use.md): Explore the computer-use topic hub.
- topic hub: [Agents](https://newruntime.com/topics/agents.md): Explore the agents topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)