Hark Handoff is a computer use agent (CUA) from Brett Adcock's Hark, announced on August 5, 2026, alongside a $700 million Series A at a $6 billion valuation. For every request it spins up a dedicated virtual computer with its own browser, file system, and terminal, then navigates any website the way a human would — clicking, scrolling, and typing through real-time decisions. It posted a 97.7 on the Online-Mind2Web benchmark, the highest ever recorded, while pricing tokens at roughly one-tenth of frontier competitors ($0.18/M input, $2.37/M output).
Key Features
- Full virtual computer per request: Each task runs in a dedicated environment with a browser, file system, and terminal — not a script against fixed sites.
- Any website, no API needed: Works with sites that have no public APIs (Target, Walmart, OpenTable, LinkedIn) by reading page structure and visual data.
- End-to-end task completion: Handoff loops observation, reasoning, and action until the task is done, predicting interactions like clicks and typing.
- Account integration: Users can connect existing accounts so the agent acts with saved addresses, preferences, and payment methods.
- 10x cheaper than frontier models: $0.18/M input and $2.37/M output tokens versus $5 and $30 for GPT-5.5-class models.
- Low latency: 0.8-second per-turn model latency with action-prediction rather than next-token prediction.
Use Cases
Who Should Use This Tool?
- Anyone drowning in browser tasks: Ordering food, booking travel, shopping, and researching across multiple sites.
- Recruiters and operators: Message candidates on LinkedIn, coordinate follow-ups, and run multi-step outreach end-to-end.
- Business teams: Placing bulk orders, returning purchases, comparing reservations across several platforms simultaneously.
Problems It Solves
- 75% of screen time in a browser: People spend over six hours a day in browsers, yet fewer than 1 in 1,000 websites have APIs.
- Multi-step web tasks are tedious: Booking flights or comparing dinner reservations across sites takes dozens of clicks.
- Assistants that can't act: Most AI assistants answer questions but can't complete real transactions on the web.
Pricing
| Plan | Price | Features |
|---|---|---|
| Beta | Waitlist | Early access signups, availability planned for end of summer |
Hark publishes token pricing rather than a subscription: $0.18 per million input tokens and $2.37 per million output tokens, with a dedicated virtual computer for each request. Public signups opened on August 5, 2026 with the platform expected later this month.
Advantages & Unique Selling Points
Compared to Competitors:
- Benchmark leader: 97.7 on Online-Mind2Web, ahead of GPT-5.4 (92.8) and Claude Opus 4.8 (84.1) in Hark's testing.
- Order-of-magnitude cost advantage: Token prices roughly one-tenth of frontier models, critical for agents that loop through dozens of actions.
- Founded by Brett Adcock: Figure AI founder's fourth startup, backed by $700M from Parkway VC, Nvidia, AMD, Qualcomm, and ARK Invest.
What Makes It Stand Out:
- Predicts the next action (a click or keystroke at a location) rather than the next token.
- "Always working, it's looping" — the agent keeps observing and acting until the job is done.
- Co-evolves model and infrastructure to handle the messy reality of the live internet.
User Reviews
Hark's launch generated strong enterprise and consumer buzz, with Adcock's demo of ordering flowers ("some of the florist's choice") going viral on social media. Coverage in VentureBeat, TechCrunch, and RuntimeWire focused on the benchmark claim and the cost advantage. The main open questions: comparisons only against prior-generation models (GPT-5.4/Opus 4.8), no published liability model for agent actions with stored payment credentials, and availability still gated behind a waitlist.
Getting Started
Quick Start Guide
- Join the beta: Sign up at hark.com — applications are being reviewed in waves.
- Connect your accounts: Link services you want Handoff to act on (saved addresses, payment methods).
- Hand off a task: Describe what you need done in plain language, from your phone or desktop.
- Review approvals: Handoff comes back only when it needs your approval or the work is finished.
Integration
- Works across any website — no API integration required.
- Connects existing accounts for personalized actions (Target, Walmart, OpenTable, LinkedIn, and more).
- Platform availability planned for end of summer, with the agent running in the cloud.
Frequently Asked Questions
Is Hark Handoff available now?
Public signups opened August 5, 2026. Hark is reviewing waitlist applications, with platform availability planned for later this month or end of summer.
How is Handoff different from a chatbot?
Chatbots answer questions. Handoff completes multi-step tasks on real websites, using its own virtual computer with a browser, file system, and terminal.
What does it cost?
Hark quotes token pricing: $0.18/M input and $2.37/M output — roughly one-tenth of GPT-5.5-class models. No subscription price has been published for the beta.
Is it safe to connect my accounts?
Hark says security and privacy are a primary focus but notes the product is a technical preview. Who is liable when an agent acts with stored payment credentials remains unspecified — a gap to watch before connecting live accounts.
Alternatives
If Hark Handoff isn't the right fit, consider these alternatives:
- Steel: Open-source browser API if you want to build your own browser agents.
- Cloudflare Computer: Browser infrastructure for agent workloads.
- Claude Cowork: Anthropic's coworker agent for organizational tasks.
Tips & Best Practices
- Start with reversible tasks: Try returns, research, and reservations before anything involving irreversible payments.
- Keep credentials scoped: Connect only the accounts a task actually needs.
- Watch for the liability question: Until Hark publishes a custody/liability model, review high-value actions before they execute.
Conclusion
Hark Handoff is the most credible bet yet that computer use agents can handle everyday web life end-to-end: a record-setting benchmark, an order-of-magnitude cost edge, and a virtual computer for every task. For anyone spending hours a day in a browser, it's the clearest glimpse of an assistant that doesn't just answer — it acts.
Comments
No comments yet. Be the first to comment!
Related Tools
Browserbase
www.browserbase.com
Browser agent infrastructure: fleets of headless browsers for AI agents, plus Search and Fetch APIs for token-efficient web context. Handles logins, CAPTCHAs, and dynamic content. Home of the Stagehand AI browser SDK.
Kitesurf
kitesurf.cloudflare.app
Cloudflare's agent-first headless browser that runs entirely in V8 isolates on Workers, using 3-7x less CPU and memory than Chromium for agentic tasks like screenshots and HTML extraction.
Daytona
daytona.io
Open-source AI agent sandbox platform: full composable environments with dedicated kernels, ~90ms creation, persistent stateful workspaces, snapshots and forking, GPU sandboxes, and computer use across Linux/Windows/macOS/Android. 72,000+ GitHub stars.
Related Insights
After I Connected Obsidian to OpenClaw, It Started Helping Me Make Decisions
Once Obsidian stopped being just a place to store notes and started working with OpenClaw, it began helping me organize context, connect information, and improve real decisions.
Stop Cramming AI Assistants into Chat Boxes: Clawdbot Picked the Wrong Battlefield
Clawdbot is convenient, but putting it inside Slack or Discord was the wrong design choice from day one. Chat tools are not for operating tasks, and AI isn't for chatting.
The Twilight of Low-Code Platforms: Why Claude Agent SDK Will Make Dify History
A deep dive from first principles of large language models on why Claude Agent SDK will replace Dify. Exploring why describing processes in natural language is more aligned with human primitive behavior patterns, and why this is the inevitable choice in the AI era.