Chapter 41: Steering a Running Agent

Steerability is a concurrency and addressing contract, not merely another prompt message.

Public-source snapshot: implementation details are examples from public Kocoro origin/main at commit 4ec6772, reviewed 2026-07-27. Treat the invariant as the lesson and re-check constants against current source.

41.1 The problem

Users change priorities while a long task is running. Treating every follow-up as a new turn creates races; silently appending it may send the final answer to the wrong inbound message. Delivery systems can also replay an already accepted message.

41.2 The design rule

Give the active run a mailbox. Drain follow-ups only at safe boundaries, attach each item to its inbound reply address, and make the model see the new instruction before selecting more tools. A delivery acknowledgement is emitted only after the terminal reply was successfully delivered, which is what lets the transport drop its replay entry — it is not a signal that the mailbox accepted the item. Use durable message identity for deduplication and a separate interrupt path for immediate cancellation.

41.3 Snapshot evidence

Source snapshot	Observation
`4ec6772`	mid-run follow-ups are drained into the active context
`4ec6772`	reply addressing is tracked per inbound item
`4ec6772`	`delivery_ack` and at-least-once replay are distinct states

The values above describe one dated implementation, not a universal contract.

41.4 Implementation checklist

Test follow-up arrival before and after a checkpoint.
Keep cancel, redirect, and ordinary context additions distinct.
Verify which inbound message receives every terminal reply.

41.5 Takeaway

Steerability is a concurrency and addressing contract, not merely another prompt message.

Read these chapters as one system: context, tool schemas, persistence, steering, time discipline, loop detection, and concurrency affect one another.

@incollection{zhang2026aiagent_en_chapter-41-steering-a-running-agent, author = {Zhang, Wayland}, title = {Chapter 41: Steering a Running Agent}, booktitle = {AI Agent Architecture: From Single Agent to Enterprise Multi-Agent Systems}, year = {2026}, url = {https://waylandz.com/ai-agent-book-en/chapter-41-steering-a-running-agent/} }