ML//agent//autoregressive control

Autoregressive control is the control of an agent by a general autoregressive model through its own output stream, with no fixed sequence of operations wired in advance (a rather strange property for anyone coming from traditional software), and it is what lets an agent choose its next step, tool or stop while it runs. It is autoregressive generation turned into a general control channel, once tokens can express tool calls and stopping decisions. Token after token, the output can represent code, an explanation, a query, a tool call, a revision of the plan, an error, another search or a decision to stop. Planning, code, tool selection and language share the same output channel, so much of the branching policy emerges during inference instead of being written as one more `if` branch.

Autoregressive control is the control of an agent by a general autoregressive model through its own output stream, with no fixed sequence of operations wired in advance (a rather strange property for anyone coming from traditional software), and it is what lets an agent choose its next step, tool or stop while it runs. It is autoregressive generation turned into a general control channel, once tokens can express tool calls and stopping decisions. Token after token, the output can represent code, an explanation, a query, a tool call, a revision of the plan, an error, another search or a decision to stop. Planning, code, tool selection and language share the same output channel, so much of the branching policy emerges during inference instead of being written as one more if branch.

That is expensive, probabilistic and hard to verify, but it explains a large part of the flexibility of agents. The agent harness does not remove that flexibility. It gives it a physics: it turns some symbolic sequences into observations and others into real actions.