iii-agentic-backend（iii Agent 后端）

Agentic Backend

Comparable to: LangGraph, CrewAI, AutoGen, Letta

Key Concepts

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

Architecture

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

Reference Implementation

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.

Common Patterns

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 state::set, state::get, state::update — 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

Adapting This Pattern

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)

Engine Configuration

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.

Pattern Boundaries

• 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.

When to Use

• 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.

Boundaries

• 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.