Hark's new web agent wants to book your flights for 10x less

AI startup Hark just launched Handoff, an autonomous web agent designed to take over the boring parts of using a browser.

Founded earlier this year by serial entrepreneur Brett Adcock, Hark landed a massive $700 million Series A round in May. Now, the company is previewing its first real product: a "computer use agent" built to navigate everyday websites end-to-end, from ordering dinner on DoorDash and booking flights on United or Delta to handling candidate messaging on LinkedIn.

Public sign-ups opened on hark.com, with a broader rollout planned for later this summer.

Virtual machines and action prediction

Hark's internal research pointed out a clear mismatch in modern computing: while people spend 75% of their daily screentime inside a web browser, less than 1 in 1,000 websites offer public APIs. That lack of connectivity has long made automated web agents brittle and impractical.

Handoff tackles the problem by spinning up an isolated virtual machine for every task, complete with its own dedicated browser, terminal, and file system. Users can link existing credentials, allowing the agent to log in, reference saved addresses, and run payments automatically.

Under the hood, Hark used a post-training pipeline featuring supervised fine-tuning and asynchronous reinforcement learning via the GRPO algorithm. Instead of predicting the next text token like a typical LLM, Handoff directly predicts the next screen action, such as clicking a specific button or typing text into a field. The startup plans to begin pre-training later this year, but currently relies on an undisclosed base model.

In a launch video, Adcock demonstrated giving the agent a voice command to assemble and order a custom bouquet from a local florist, complete with vague parameters like "florist's choice." He also claimed he now runs all of his recruiting workflows through Handoff.

Impressive benchmarks, with notable asterisks

Hark claims Handoff scored a record 97.7 on Online-Mind2Web, an established web-agent benchmark with a human-evaluated leaderboard. That edges out OpenAI's GPT 5.4 at 92.8, Anthropic's Claude Opus 4.8 at 84.1, and Google's Gemini 2.5 Pro at 69.

Even more striking is the cost. Hark quotes Handoff at $0.18 per million input tokens and $2.37 per million output tokens, alongside an average turn latency of 0.8 seconds. By comparison, GPT 5.5 runs at $5 per million input and $30 per million output, with Hark measuring its latency at over 6 seconds.

Still, the comparisons come with heavy caveats. Hark measured its agent against last-generation systems, excluding frontier updates like GPT-5.6, Claude Opus 5, and open-source models like DeepSeek V4 or Qwen3.8-Max. On the broader OSWorld 2.0 benchmark, Opus 5 reportedly hits around 70.6%, far above the 55.7% scored by Opus 4.8.

Furthermore, competitor latencies were evaluated using Hark's internal testing setup at maximum reasoning settings, and on WebTailBench v2, GPT 5.5 actually beat Handoff with a score of 72.3 to 68.6. Two of the three evaluations provided were run inside Hark's own test harness and judged by an internal LLM.

Why it matters

If web agents can reliably complete complex chores across uncooperative websites, they will fundamentally change everyday internet use. Booking complicated flights or returning retail items on Target and Walmart without touching a keyboard sounds great. That said, handing over active credentials and payment cards to an unreleased agent carries serious privacy and security risks. We think it is worth joining the waitlist, but users should wait for independent testing before letting it touch financial accounts.

Hark is jumping into a crowded pool that includes tech giants like Google and Anthropic, alongside funded browser startups like Browser Use, Polar, and Aside. Hark's smart move is focusing aggressively on post-training efficiency and cost, offering high-speed browser actions at a tenth of frontier prices.

However, building on an external base model leaves Hark vulnerable. If Anthropic or OpenAI further slash prices or bake specialized browser execution directly into their frontier APIs, a post-trained wrapper can lose its competitive edge overnight. Hark's real test will be whether its specialized virtual environment can deliver reliability that general-purpose foundation models cannot match.