Offloop launches AI Agent Workspace for recurrent work

Offloop has opened its agent workspace in private beta, providing invite-based access to a system that maintains recurring teamwork within persistent Channels instead of restarting it with every prompt. The product is designed for small teams managing launches, customer follow-up, research, feedback, growth, and operations, where progress often stalls not due to hard problems but due to follow-through that no one has time to chase.

The unit of work is the Channel. People, agents, messages, files, schedules, signals, tool progress, and finished artifacts are all located in the same place, allowing an agent picking up a task to inherit the context, owners, and permissions that preceded it. Schedules and incoming signals can independently activate an agent, transforming a one-off request into a loop that continues running. In one setup demonstrated by the company, an agent is granted access to a Google Workspace inbox and subsequently activates the Channel with a summary whenever a matching email arrives. Actions remain visible throughout, and individuals can inspect, redirect, approve, or stop a run at any point.

Coordinating several agents within one Channel is managed by D1, a dispatcher model developed by Offloop for this purpose. With multiple agents sharing a workspace, a decision-maker is needed to determine which agent acts at each step and to respond when context shifts or new updates occur, and D1 fulfills this role. Teams can connect their own AI subscription account or utilize inference provided by Offloop. Connectors currently include Notion, GitHub, GitLab, Sentry, Supabase, Vercel, Railway, Cursor, Codex, and ChatGPT, with a macOS desktop app accompanying the workspace.

Offloop is also quantifying the multi-agent framework. According to figures released by the company, its multi-agent system scores 84.9 percent on GDPval and 67.2 percent on JobBench, surpassing the frontier multi-agent baselines it tested against, at approximately a fifth of the cost per task. GDPval is OpenAI's benchmark of real-world economically valuable work spanning 44 occupations across nine sectors, while JobBench, published by researchers at the University of Washington, evaluates agents on the tasks professionals wish to delegate. Offloop presents the result as the first multi-agent system to demonstrate its value on real-world benchmarks.

The company behind the product is Intelligence Software, Inc., and Offloop states that the team previously developed general-purpose agents at scale and recognized where a single agent reaches its limits. Access is available across three tiers: an invite-only Operator plan for individuals converting recurring work into agent-assisted Channels, a Team Pilot plan that adds shared Channels, team memory, connector setup, and priority onboarding, and an enterprise tier that includes custom connectors, SSO planning, and an audit-friendly runtime. Invite codes for the beta are being distributed through the Offloop site.