How we built teams of AI agents that don't loop

Most demos of multi-agent systems feature a few agents chatting in a shared space until something useful emerges. However, this approach is problematic when applied to products, as it leads to loops, duplicate work, agents answering unnecessary questions, and escalating token costs with each round. AWFlow, a browser extension for automation, recently launched agents and teams of agents to address these issues.

Each agent in AWFlow comes equipped with its own model, instructions, memory, skills, and rules. The coordination runtime is the key component that makes teams of agents usable. In the post, the focus is on the coordination runtime, which ensures teams of agents behave effectively and efficiently.

A team of people operates with a defined way of working, a clear map of who hands work to whom, and set budgets. In contrast, a free-for-all coordination, typical in most demos, results in loops, duplicate work, and growing token bills. AWFlow's coordination runtime operates on a team snapshot, which includes members, roles, the way of working, the handoff map, rules, and limits. When a thread starts, the runtime uses this snapshot to make decisions. Changes to the team can't silently alter the rules during a conversation.

AWFlow presents three ways of working for teams. The first is the classic manager shape, where a lead merges the parts into one answer. This method works well for open-ended goals like "plan a weekend in Lisbon". The second method, the pipeline, removes planning altogether and has members running in a fixed order, receiving the previous result.

Each member uses its own model, skills, and rules and returns a result back to the agent that asked. The third and last method, the open table, is for when a human is steering. Handoffs follow a map, and the tool only exists when the map specifies someone to hand to. Agents that can't be handed to don't have a tool, making the process more cost-effective and reliable.