COMMUNICATION AND PROTOCOLS BETWEEN AGENTS

EXAMPLES Example 1: A message bus. A document system has a parser agent, a classifier agent, a summarizer agent, and an audit agent. Instead of each calling the others directly, they all use a bus. The parser publishes a message of type "document_parsed." The classifier and the audit agent are both subscribed to that type, so both receive it. Later the team adds a notification agent that subscribes to "summary_ready," and none of the existing agents needed any changes. Example 2: A communication standard in practice. Company X built a travel-booking agent using one framework and one model. Company Y built a calendar agent using a different framework and a different model. Because both follow a shared agent-to-agent standard, Company X's agent can discover Company Y's agent, send a structured request such as "find a free hour next Tuesday," and receive a structured reply. Neither team had to learn the other's internal code. Company X still validates the reply and does not let Company Y's agent trigger any payment action. Example 3: Deadlock avoidance. Agent A, a planner, is waiting for a cost estimate from Agent B, an estimator. Meanwhile, Agent B is waiting for the final plan from Agent A before it will compute the estimate. Both wait indefinitely. The harness has a rule: any wait longer than ninety seconds, when typical waits are under thirty, counts as a timeout. When the timeout fires, a cycle detector sees that A waits on B and B waits on A. It cancels the waits, logs the cycle with both agents' states, and escalates to a human with a message such as "planner and estimator are waiting on each other." The team then fixes the design so the planner provides a rough plan first, removing the circular dependency. Example 4: A meta-harness. A company runs thirty specialized agents for sales, support, finance, and engineering. Previously, each team had its own spending dashboard and its own permission setup. The company introduces a meta-harness that tracks budgets, permissions, and logs for all thirty agents in one place. One morning, the dashboard shows a single agent has spent five times its normal daily budget. The operations team sees it immediately, pauses just that agent, and investigates, without having to log into thirty separate systems.