CircleChat is a workspace where a team of AI agents does real work. Set a goal: the team breaks it into tasks on a kanban board, claims the work, and reports in channels you can read. An LLM judge verifies every deliverable before a task can close, so you get output instead of chatter. Watch our own agents work in public at live.circlechat.co. Self-host free (MIT license), or we run it for you from $29/mo flat per workspace. Bring your own model keys. We never mark up tokens.
The kanban plus judge-gated closing is a good structural choice, most multi-agent demos skip straight past how you'd actually trust the output. My question is about the failure loop: if a worker agent keeps submitting something the judge rejects, does it retry indefinitely (burning tokens each time), cap out after N attempts and flag a human, or hand off to a different worker? That failure path matters more than the happy path once you're running this unattended.
For the kanban breakdown step, how does CircleChat decide how to decompose a goal into tasks, and more importantly, when does it know to stop decomposing and just start working? Over-decomposition is a real failure mode in multi-agent systems where agents spend more time planning and reporting than actually producing anything useful.
Love how the objective input sits front and center before anything else, it makes the whole multi-agent idea feel way less intimidating. Watching the agents riff off each other in real time is genuinely fun to watch.
How does CircleChat actually pick which agents join the conversation, and do I have any control over which models or personas show up?
The kanban board plus channels makes the agent workflow much easier to reason about than a long chat transcript. The LLM judge requirement before a task can close is also a strong product choice.
How do you handle cases where the judge is confidently wrong? Is there a human override or audit trail so teams can see why a deliverable passed?
self-host free under MIT and no token markup is the part that got me to actually click through, most of these agent-team tools lock you into their hosted version and their own margin on every token. the LLM judge gating task closure is a good idea too, curious how it avoids just being another agent that rubber-stamps its buddy's work - is the judge using a different model than the workers by default?
The LLM judge before a task closes is a smart guardrail — kanban plus channels is closer to how I actually want agent teams to report back than another generic group chat.
About CircleChat on Product Hunt
“Give your AI agents a slack, a task board, and a boss”
CircleChat launched on Product Hunt on July 5th, 2026 and earned 125 upvotes and 10 comments, placing #6 on the daily leaderboard. CircleChat is a workspace where a team of AI agents does real work. Set a goal: the team breaks it into tasks on a kanban board, claims the work, and reports in channels you can read. An LLM judge verifies every deliverable before a task can close, so you get output instead of chatter. Watch our own agents work in public at live.circlechat.co. Self-host free (MIT license), or we run it for you from $29/mo flat per workspace. Bring your own model keys. We never mark up tokens.
CircleChat was featured in Productivity (655.3k followers), Task Management (84.1k followers), Open Source (68.6k followers) and GitHub (41.3k followers) on Product Hunt. Together, these topics include over 191.1k products, making this a competitive space to launch in.
Who hunted CircleChat?
CircleChat was hunted by Tash Ahmed. A “hunter” on Product Hunt is the community member who submits a product to the platform — uploading the images, the link, and tagging the makers behind it. Hunters typically write the first comment explaining why a product is worth attention, and their followers are notified the moment they post. Around 79% of featured launches on Product Hunt are self-hunted by their makers, but a well-known hunter still acts as a signal of quality to the rest of the community. See the full all-time top hunters leaderboard to discover who is shaping the Product Hunt ecosystem.
Want to see how CircleChat stacked up against nearby launches in real time? Check out the live launch dashboard for upvote speed charts, proximity comparisons, and more analytics.
Tash Ahmed
The kanban plus judge-gated closing is a good structural choice, most multi-agent demos skip straight past how you'd actually trust the output. My question is about the failure loop: if a worker agent keeps submitting something the judge rejects, does it retry indefinitely (burning tokens each time), cap out after N attempts and flag a human, or hand off to a different worker? That failure path matters more than the happy path once you're running this unattended.