Supervisor
A multi-agent architecture pattern in which a supervisor agent coordinates the workflow of multiple specialized agents, such as researchers and writers.
Timeline (last 90 days)
In the 240-question test, ReAct's accuracy was 59.6%, Plan-execute's 61.7%, Supervisor's 65.0%, and Draft-verify's 63.7%.
ReAct consumed an average of 4,663 tokens per question, Plan-execute 10,535 tokens, Supervisor 7,494 tokens, and Draft-verify 4,905 tokens.
For simple task dispatch, tool calling is enough; only scenarios that truly require multi-round negotiation and dynamic planning are worth bringing in a full Supervisor.
Passing only the latest user message plus a short summary (rather than the entire history) to the router, and attaching a pre_model_hook to the Supervisor, can streamline its input and reduce Token consumption.
The Supervisor runs on every hop, and each time it wakes up it must re-read the shared message list, which contains the user's input for each round, each of the Supervisor's own previous routing calls, the transfer message pair produced by each handoff/return, and the last message of each worker.
The root cause of the non-linear growth of Token consumption in multi-agent systems is the number of calls multiplied by the ever-expanding context, not the cost of a single call.
Handling unambiguous intents directly with code before they enter the LLM can reduce the Supervisor's Token consumption.
In a Supervisor architecture, placing a changing timestamp or dynamic few-shot examples at the top of the prompt breaks the prefix cache, multiplying routing costs.
Common orchestration patterns for multi-Agent mainly include the Supervisor pattern, the Network/Swarm pattern (direct handoff between Agents, with no central node) and the Hierarchical pattern (a supervisor nested under another supervisor, suitable for large-scale systems).
A Reddit developer's implementation pain point: the Supervisor only classifies intent on the first message of a session and subsequently routes to the same Agent, causing topic-switching failures and the absence of multi-Agent collaboration.
2 more timeline events
All Facts (13)
传统进程管理工具如Supervisor和PM2无法区分AI Agent的语义状态，不能分辨'任务正在思考'与'任务已经卡死'
90%Unverified当前Agent编排领域存在抽象层级错配的两极分化格局：重量级方案如LangGraph、Temporal、AWS Step Functions抽象层级过高，轻量级工具如PM2、Supervisor抽象层级过低
80%UnverifiedReAct consumed an average of 4,663 tokens per question, Plan-execute 10,535 tokens, Supervisor 7,494 tokens, and Draft-verify 4,905 tokens.
50%UnverifiedIn the 240-question test, ReAct's accuracy was 59.6%, Plan-execute's 61.7%, Supervisor's 65.0%, and Draft-verify's 63.7%.
50%UnverifiedHandling unambiguous intents directly with code before they enter the LLM can reduce the Supervisor's Token consumption.
50%UnverifiedIn a Supervisor architecture, placing a changing timestamp or dynamic few-shot examples at the top of the prompt breaks the prefix cache, multiplying routing costs.
50%UnverifiedPassing only the latest user message plus a short summary (rather than the entire history) to the router, and attaching a pre_model_hook to the Supervisor, can streamline its input and reduce Token consumption.
50%UnverifiedFor simple task dispatch, tool calling is enough; only scenarios that truly require multi-round negotiation and dynamic planning are worth bringing in a full Supervisor.
50%UnverifiedThe Supervisor runs on every hop, and each time it wakes up it must re-read the shared message list, which contains the user's input for each round, each of the Supervisor's own previous routing calls, the transfer message pair produced by each handoff/return, and the last message of each worker.
50%UnverifiedThe root cause of the non-linear growth of Token consumption in multi-agent systems is the number of calls multiplied by the ever-expanding context, not the cost of a single call.
50%UnverifiedCommon orchestration patterns for multi-Agent mainly include the Supervisor pattern, the Network/Swarm pattern (direct handoff between Agents, with no central node) and the Hierarchical pattern (a supervisor nested under another supervisor, suitable for large-scale systems).
50%UnverifiedA more robust approach is to have the Supervisor dynamically construct a task DAG at every turn of the conversation, rather than binding an Agent in a one-time fashion.
50%UnverifiedErlang/OTP的Supervisor模块提供one_for_one、one_for_all、rest_for_one三种子进程崩溃恢复策略，体现「let it crash」哲学
50%