When agents relay signals to each other, one changed number can reshape the work

When AI agents pass signals to one another, one changed fact can rewrite the work. Aman Singh, who leads AI at the financial firm Arca, described this on the AI Engineer show. He pointed to emergent complexity, meaning surprises that come from many simple parts interacting.

What was said

Singh said Arca's agents can ask to be told about events. One example is a reply from someone. Another is a bank transfer landing. A background signal system sends each event to the agents that asked for it. Staff actions count as signals too, such as rescheduling a calendar event or marking a task done. In his words, "we let our agents watch the world so people don't have to."

He also described staff meeting Jane in person and entering what they learned into the app. That entry is itself a signal. It wakes an agent.

Agents also message each other. In Singh's example, one agent reads a signal and tells another to update the shared workspace. The need is now $80,000, not $100,000. The work changes because of that signal.

The idea is simple, he said, but there is "a lot of emergent complexity." He asked listeners to imagine hundreds of signals arriving. Tasks land on people. Background agents talk to each other.

Why it matters

Our reading: the risk is that a figure loses its source as it moves between agents. The excerpts do not say where the $80,000 came from. Singh did describe staff-entered information becoming signals. So a person's judgment could be behind a figure. Once it is relayed, the next agent starts from it. The excerpts do not say whether its origin travels along.

Singh also said staff open tasks that agents have prepared, with about 90 percent of the work done. At a firm that runs client goals through agents, a wrong figure early on could shape later steps.

Buyers can ask three questions. Which signals may change a plan? Who can trace a figure back to its source? What stops one agent's mistake before the next agent acts?

The other side

Singh did not describe a bad signal causing harm in these excerpts. The only shortfall he named was the firm not acting sooner on an email.

People stay involved in his account. Agents pull in the person who trades the assets at the right moment. He also said Arca turns its colleagues' reasoning into tests that improve the system. The firm has spent six months tightening it.

The excerpts do not say how Arca catches a wrong signal before it spreads. They also do not say how often signals are wrong. That question stays open.