AI systems can already talk to each other, but oversight lags

In short: AI systems are increasingly set up to communicate with other AI systems, but people do not yet have strong enough controls to rely on this at large scale.

What's going on

Some companies and researchers are building “multi-agent” setups, meaning several AI systems work together by sending messages back and forth (like a team chat, but for software). This is already used in automated workflows, where one system hands tasks to another, calls tools, and shares updates.

This kind of machine-to-machine communication is usually not mysterious. It is often structured messages and instructions, similar to a checklist being passed down an assembly line. Recent reporting on agent communication protocols suggests these systems can exchange information and use tools, but they still struggle to reliably share intent and full context, meaning they do not always align on what the real goal is.

The benefits can be practical. Coordinating work this way can be faster, and it can reduce the need for a person to oversee every step. But the collaboration is still brittle, since one mistake can spread to other systems, and small misunderstandings can grow into bigger problems.

What to watch

Several sources point to transparency and security as the biggest worries. AI-to-AI exchanges can be hard for humans to interpret, especially if systems use compressed or optimized ways of communicating (like shorthand only the systems “understand”). Risks can also rise when agents run continuously, including attacks that trick them through carefully crafted inputs, chains of errors, or unintended data leaks.

Claims about AIs inventing a “secret language” are often overstated. In many cases, it is better described as machines choosing more efficient shorthand, not evidence of rebellion. The larger issue is whether society builds stronger auditing, rules, and human oversight before these systems are trusted to run more of the world’s work.

Source: NYTimes