agentmanager
由 nonightwatch 提供
Communication简介
This file is a concise integration contract for AI tool callers and gateway implementers.
描述
该技能的功能
Agent Manager is an orchestration kernel that receives plans from external AI systems, then validates, schedules, and executes the tools within those plans. It enforces token budgets, tracks per-task costs, and stores replayable telemetry for every run.
It separates plan authoring from execution infrastructure, so any external planner can submit plans without implementing its own scheduling, budget enforcement, or streaming telemetry.
何时使用
- Running multi-step AI agent workflows with enforced cost budgets
- Streaming real-time task execution events to a monitoring dashboard
- Replaying a failed run to debug the exact tool call sequence
- Injecting deterministic result overrides into a live task mid-run
- Auditing per-task LLM costs across multiple tenants or workloads
示例工作流
User asks: execute a two-task plan where task A fetches data and task B summarizes it
- 1
Call GET /v1/capabilities to confirm available providers
- 2
Submit plan via POST /v1/plan to validate structure and get recommendations
- 3
Start the run with POST /v1/run and receive a run ID
- 4
Stream progress with GET /v1/run/:id/stream, reading task_start and tool_call_end events
- 5
Retrieve final results and cost breakdown from GET /v1/run/:id
Completed run record with per-task cost attribution and a replayable event log
安装
clawhub install nonightwatch/agentmanager运行 CLI 命令,或把这段提示词粘贴到 Claude Code、Cursor、Codex 或你在用的任意编程智能体里。