iii-agentic-backend

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Agentic Backend

Agent化后端

Comparable to: LangGraph, CrewAI, AutoGen, Letta

同类工具：LangGraph、CrewAI、AutoGen、Letta

Key Concepts

核心概念

Use the concepts below when they fit the task. Not every agentic workflow needs all of them.

Each agent is a registered function with a single responsibility
Agents communicate via named queues (ordered handoffs) and shared state (accumulated context)
Approval gates are explicit checks in the producing agent before enqueuing the next step
An HTTP trigger provides the entry point; agents chain from there
Pubsub broadcasts completion events for downstream listeners

请根据任务需求选用以下概念。并非所有Agent化工作流都需要用到全部概念。

每个Agent都是一个具有单一职责的已注册函数
Agent通过命名队列（有序移交）和共享状态（累积上下文）进行通信
审批关卡是生产Agent在将下一步加入队列前执行的显式检查
HTTP触发器作为入口点，Agent从此处开始链式执行
Pubsub（发布订阅）向下游监听者广播完成事件

Architecture

架构

text

HTTP request
  → Enqueue(agent-tasks) → Agent 1 (researcher) → writes state
    → Enqueue(agent-tasks) → Agent 2 (critic) → reads/updates state
      → explicit approval check (is-approved?)
        → Enqueue(agent-tasks) → Agent 3 (synthesizer) → final state update
          → publish(research.complete)

text

HTTP request
  → Enqueue(agent-tasks) → Agent 1 (researcher) → writes state
    → Enqueue(agent-tasks) → Agent 2 (critic) → reads/updates state
      → explicit approval check (is-approved?)
        → Enqueue(agent-tasks) → Agent 3 (synthesizer) → final state update
          → publish(research.complete)

iii Primitives Used

使用的iii原语

Primitive	Purpose
`registerWorker`	Initialize the worker and connect to iii
`registerFunction`	Define each agent
`trigger({ function_id: 'state::set/get/update', payload })`	Shared context between agents
`trigger({ ..., action: TriggerAction.Enqueue({ queue }) })`	Async handoff between agents via named queue
`trigger({ function_id, payload })`	Explicit condition check before enqueuing
`trigger({ function_id: 'publish', payload, action: TriggerAction.Void() })`	Broadcast completion to any listeners
`registerTrigger({ type: 'http' })`	Entry point

原语	用途
`registerWorker`	初始化Worker并连接到iii引擎
`registerFunction`	定义每个Agent
`trigger({ function_id: 'state::set/get/update', payload })`	实现Agent间的共享上下文
`trigger({ ..., action: TriggerAction.Enqueue({ queue }) })`	通过命名队列实现Agent间的异步移交
`trigger({ function_id, payload })`	加入队列前执行显式条件检查
`trigger({ function_id: 'publish', payload, action: TriggerAction.Void() })`	向所有监听者广播完成事件
`registerTrigger({ type: 'http' })`	作为入口点

Reference Implementation

参考实现

See ../references/agentic-backend.js for the full working example — a multi-agent research pipeline where a researcher gathers findings, a critic reviews them, and a synthesizer produces a final report.

完整可运行示例请查看../references/agentic-backend.js —— 这是一个多Agent研究流水线，其中研究员Agent收集研究结果，评审员Agent进行审核，合成员Agent生成最终报告。

Common Patterns

常见模式

Code using this pattern commonly includes, when relevant:

```
registerWorker(url, { workerName })
```
— worker initialization

trigger({ function_id, payload, action: TriggerAction.Enqueue({ queue }) })

— async handoff between agents

trigger({ function_id: 'state::set/get/update', payload: { scope, key } })

— shared context between agents

Explicit condition check via

await iii.trigger({ function_id: 'condition-fn', payload })

before enqueuing next agent

trigger({ function_id: 'publish', payload: { topic, data }, action: TriggerAction.Void() })

— completion broadcast

Each agent as its own
```
registerFunction
```
with
```
agents::
```
prefix IDs
```
const logger = new Logger()
```
— structured logging per agent

使用此模式的代码通常包含以下相关内容：

```
registerWorker(url, { workerName })
```
—— Worker初始化

trigger({ function_id, payload, action: TriggerAction.Enqueue({ queue }) })

—— Agent间的异步移交

trigger({ function_id: 'state::set/get/update', payload: { scope, key } })

—— Agent间的共享上下文

在加入下一个Agent队列前，通过
```
await iii.trigger({ function_id: 'condition-fn', payload })
```
执行显式条件检查

trigger({ function_id: 'publish', payload: { topic, data }, action: TriggerAction.Void() })

—— 完成事件广播

每个Agent都是带有
```
agents::
```
前缀ID的独立
```
registerFunction
```
```
const logger = new Logger()
```
—— 每个Agent的结构化日志

Adapting This Pattern

适配此模式

Use the adaptations below when they apply to the task.

Replace simulated logic in each agent with real work (API calls, LLM inference, etc.)
Add more agents by registering functions and enqueuing to them with
```
TriggerAction.Enqueue({ queue })
```
For approval gates, call a condition function explicitly before enqueuing the next agent
Define queue configs (retries, concurrency) in
```
iii-config.yaml
```
under
```
queue_configs
```
State scope should be named for your domain (e.g.
```
research-tasks
```
,
```
support-tickets
```
)
```
functionId
```
segments should reflect your agent hierarchy (e.g.
```
agents::researcher
```
,
```
agents::critic
```
)

请根据任务需求选用以下适配方式。

将每个Agent中的模拟逻辑替换为实际工作（API调用、LLM推理等）
通过注册函数并使用
```
TriggerAction.Enqueue({ queue })
```
将任务加入队列，添加更多Agent
对于审批关卡，在加入下一个Agent队列前显式调用条件函数
在
```
iii-config.yaml
```
的
```
queue_configs
```
下定义队列配置（重试、并发数）
状态作用域应根据你的业务领域命名（例如
```
research-tasks
```
、
```
support-tickets
```
）
```
functionId
```
分段应体现你的Agent层级结构（例如
```
agents::researcher
```
、
```
agents::critic
```
）

Engine Configuration

引擎配置

Named queues for agent handoffs are declared in iii-config.yaml under

queue_configs

. See ../references/iii-config.yaml for the full annotated config reference.

Agent移交使用的命名队列需在

iii-config.yaml

的

queue_configs

中声明。完整带注释的配置参考请查看../references/iii-config.yaml。

Pattern Boundaries

模式边界

If a request is about adapting existing HTTP endpoints into
```
registerFunction
```
(including prompts asking for
```
{ path, id }
```
endpoint maps + loops), prefer
```
iii-http-invoked-functions
```
.
Stay with
```
iii-agentic-backend
```
when the primary problem is multi-agent orchestration, queue handoffs, approval gates, and shared context.

如果需求是将现有HTTP端点适配为
```
registerFunction
```
（包括要求提供
```
{ path, id }
```
端点映射及循环的请求），请优先使用
```
iii-http-invoked-functions
```
模式。
当核心需求是多Agent编排、队列移交、审批关卡和共享上下文时，使用
```
iii-agentic-backend
```
模式。

When to Use

使用场景

Use this skill when the task is primarily about
```
iii-agentic-backend
```
in the iii engine.
Triggers when the request directly asks for this pattern or an equivalent implementation.

当任务主要涉及iii引擎中的
```
iii-agentic-backend
```
时，使用此技能。
当请求直接要求实现此模式或等效方案时触发。

Boundaries

使用限制

Never use this skill as a generic fallback for unrelated tasks.
You must not apply this skill when a more specific iii skill is a better fit.
Always verify environment and safety constraints before applying examples from this skill.

切勿将此技能作为无关任务的通用 fallback。
当有更适合的特定iii技能时，禁止使用此技能。
在应用此技能中的示例前，务必验证环境和安全约束。