What You Need to Know
ACP thread agents let you assign a dedicated agent runtime to each thread or conversation session. Instead of a single monolithic agent handling all conversations, you can spawn specialized agents that are purpose-built for specific tasks. A support thread gets a support agent, a code review thread gets a dev agent, and each operates independently.
The spawn/send dispatch model is straightforward. When a new thread starts, the orchestrator spawns an ACP agent with a specific config (model, skills, system prompt, constraints). Messages in that thread are dispatched to the spawned agent via send. The agent processes the message within its own context and returns the response through the same channel.
Lifecycle controls give you fine-grained management over agent instances. You can monitor active agents, check their resource usage, pause or resume them, and terminate agents that are no longer needed. This prevents runaway resource consumption in high-traffic environments where many threads are active simultaneously.
The practical benefit is isolation without infrastructure complexity. Previously, running different agent configurations for different conversations required multiple gateway instances or complex routing rules. ACP thread agents handle this natively — one gateway, many specialized agents, each with their own context window, tool permissions, and behavioral constraints.
This pattern is especially powerful for multi-user team deployments. Each team member's conversation gets an agent instance that respects their specific permissions and context. When the conversation ends, the agent's session is compacted, archived, and the runtime resources are reclaimed. No manual cleanup, no context cross-contamination.