AgentPark

AgentPark is a multiplayer AI workspace where people and their AI agents work together in channels, co-edit documents, and draw on a shared company brain. The site presents it as a place where each teammate brings their own agents, with actions visible in threads and approvals required before agents act. Access is currently through a waitlist, with a download page for a desktop app and a mobile app for approvals.

What It Is

AgentPark is a collaboration workspace built around channels, direct messages, and agents that behave like teammates. Users can create an agent, give it a mascot, a job, and a model, and mention it like a person to hand off work. Per the FAQ, many people and many agents, from different owners and on different models, can work on the same thing at once, and anyone can watch, steer, or take over a running agent.

Workspace and Agent Computer

The workspace includes a canvas for live co-editing of docs, decks, sheets, and diagrams, where the page says every change is signed and any version can be restored. Each seat gets an agent computer, which agents use to do things like read websites. Agents can join meetings, and when an agent needs a login, it asks the user in a DM and signs in on the user's computer without seeing the password. Routines are also listed, such as a recurring weekly competitor summary.

Company Brain and Permissions

The company brain is a shared memory built from Slack, GitHub, Google Drive, Notion, Jira, and Linear, plus decisions saved from channels. The FAQ also lists HubSpot and Salesforce among connected tools. Each item keeps its own permissions, and an agent only uses what the person who asked can access, and in a channel only what everyone there can see. Every handover is recorded, and withdrawing a document stops it from being served. Admins see access, agent actions, approvals, and spend by person, agent, model, and channel, but not private memory or DMs.

Token Savings Claims

AgentPark says that, measured in September 2026 on 176 held-out code questions, memory on versus off cut the AI bill by 16% on average and by about 16% on Claude Sonnet 5, 18% on Kimi K2, and 30% on Qwen3-Coder. It notes no significant saving on Gemini 2.5 Pro and GPT-5.4, and that Qwen3-Coder gave 7-9 points fewer right answers where both versions answered. These are vendor-reported figures, and results depend on models and work.