Supervisor
A multi-agent architecture pattern in which a supervisor agent coordinates the workflow of multiple specialized agents, such as researchers and writers.
Core Facts
Timeline (last 90 days)
OpenShell uses the operating system kernel to confine agents in a sandbox, restricting the files they can access and the processes they can start, and completely blocking their network, with outbound requests checked against policy by a Supervisor outside the sandbox.
Common collaboration patterns in multi-agent systems include pipeline (sequential dependency), parallel (independent subtasks), and debate (multiple agents discussing to reach consensus).
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.
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%.
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.
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.
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.
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.
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.
4 more timeline events
All Facts (15)
Common collaboration patterns in multi-agent systems include pipeline (sequential dependency), parallel (independent subtasks), and debate (multiple agents discussing to reach consensus).
75%Unverified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%UnverifiedOpenShell uses the operating system kernel to confine agents in a sandbox, restricting the files they can access and the processes they can start, and completely blocking their network, with outbound requests checked against policy by a Supervisor outside the sandbox.
50%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%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%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%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%