Clark

An AI coworker with its own cloud computer

Productivity

Developer Tools

Artificial Intelligence

Visit WebsiteSee on Product HuntTwitter ⧉

We stress-tested Clark on original mathematics research — specifically, an open gap in our Riemann Hypothesis program at MathLab (mathlab.quantiterate.com), a proof verification platform with an adversarial War Room.

We gave Clark a five-step forward path and told him to implement it. He produced two compilable LaTeX files: an unconditional lemma with correct proofs and a conditional strip theorem with five named hypotheses. When we critiqued structural defects in v1.0, he absorbed every point and produced a tighter v1.1 with a changelog. Rating: 7–8/10.

Your three questions:

1. First task: we went straight to the hardest thing we could think of. Most honest test.

2. "Own cloud computer" comes through — getting back compilable LaTeX with correct theorem numbering and cross-references is qualitatively different from chat.

3. Where we'd hesitate: autonomous long-horizon work. Clark is an excellent decomposer — he factors hard problems into clean sub-problems. But when the difficulty is a tension that doesn't decompose further, he'll keep refining the scaffold without resolving it. The human critique loop is what made him improve. Infrastructure, not engine. We'll use him again.

giving the agent its own cloud computer is the right bet imo 👏 the async close-the-tab flow is nice. hows reliability on longer tasks?

Congrats on the launch — #2 on the day is well earned. The under-discussed part IMO is the OpenAI-compatible API: an embeddable agent means other agents can delegate to Clark, and agent-to-agent work is where a lot of this space is heading. How does the API handle long-running tasks — does a caller get a job handle to poll, streamed progress, or a webhook when the artifact is ready?

the async model is the right call. most AI tools force you to babysit the process in real-time which defeats the purpose. curious how you handle the verification step though — when Clark finishes a research task, how does the user know the sources are solid without re-doing the work themselves?

This looks cool! Was wondering what the difference is between this and /loop on claude code? Can't cc agents run autonomously already?

Love that it actually hands you back evidence with the work, not just a wall of text. Closing the tab and coming back to a finished spreadsheet feels like cheating.

Love that Clark can fan work out to parallel specialists and hand back artifacts with sources. One thing I'd love is a simple "trust but verify" mode where I can quickly diff what Clark did against my own expectations before it writes to a repo or ships a spreadsheet, so I can catch small errors without babysitting the run.

the scheduled monitor use case is the one I'd want most out of this. when it re-runs on a timer, does it diff against the last artifact and surface only what changed, or hand back a fresh full artifact every time and leave the comparison to you?

the scheduled monitoring runs are the part I'd worry about most - if a recurring job silently starts failing (site changed, credentials expired) does it flag that as a failure or just quietly hand back a stale/empty artifact next run?

the scheduled monitoring use case is the interesting one - when a monitor job runs repeatedly, does it diff against the last run so you only get pinged on what changed, or does it hand back the full state fresh each time and you compare yourself?

How does the fan out to parallel specialists actually work, does Clark decide on its own how to split a task or do you have to define the sub agents ahead of time?

Moving from “answer my prompt” to “complete the work and return evidence” is the interesting shift here. For long-running tasks, how does Clark preserve the reasoning and context behind intermediate decisions so users can inspect more than just the final artifact?

the sandboxing/state questions above cover the trust side well. curious about the fan-out mechanic specifically - when Clark splits a task across parallel specialists, do they each get an isolated environment that gets merged into the final artifact, or do they share the same cloud computer/workspace while running concurrently? asking because shared-state parallelism is where I'd expect the weird bugs to show up first

an AI coworker with its own cloud computer is a wild framing. curious how you handle the case where it needs to install something or hit a paywall/captcha mid-task - does it just stall out and ping you, or does it have some way to work around that on its own?

Congrats on the launch! How does Clark compare to Hermes Agent deployed on a cloud computer? Does the workflow / tool-chain is built to be more complete towards problem-solving?

"Its own cloud computer" is a meaningfully different architecture than most AI-coworker tools that just call APIs — giving it a persistent environment changes what it can actually do end-to-end. Curious how you're handling security/sandboxing for that computer, especially once it's doing real multi-step tasks unsupervised. That's usually the part that keeps teams from trusting autonomous agents with anything consequential.

Congrats on the launch! How does Clark deal with existing internal systems that have messy or outdated APIs?

Congrats on the launch @stanislav_kirdey. How does Clark handle changes after an app is generated? Can it keep updating the app as requirements evolve?