ai-agent

지속적 메모리를 갖춘 AI 에이전트 백엔드 구축: 대화당 하나의 Rivet Actor, 대기열 메시지 처리, 실시간 이벤트로 스트리밍되는 LLM 응답.

npx skills add https://github.com/rivet-dev/skills --skill ai-agent

AI Agent

IMPORTANT: Before doing anything, you MUST read BASE_SKILL.md in this skill's directory. It contains essential guidance on debugging, error handling, state management, deployment, and project setup. Those rules and patterns apply to all RivetKit work. Everything below assumes you have already read and understood it.

Working Examples

If you need a reference implementation, read the raw working example code in these templates:

ai-agent

Patterns for building AI agent backends with RivetKit, where each conversation is one Rivet Actor that owns its memory, its message queue, and its streaming output.

Starter Code

Start with one of the working examples on GitHub and adapt it. The sections below describe the flagship ai-agent example unless a variant is called out explicitly.

Variant	Starter Code	Use When
Queue-driven AI SDK agent	GitHub	You want a streaming chat agent where each conversation keeps its own persistent memory and processes one message at a time.
Sandbox coding agent	GitHub	The agent should run a coding agent (Codex by default) inside an isolated sandbox via Docker, Daytona, or E2B.
Durable streams agent (experimental)	GitHub	You want replayable, restart-safe prompt and response delivery through durable streams instead of actor state and events.

Conversation Memory

Use one actor per conversation, keyed by a conversation or agent id (see Actor Keys). The agent actor's persistent state is the conversation memory: in the ai-agent example, messages and status live in JSON actor state and survive sleep and restarts with no external database. Every model call rebuilds the prompt from c.state.messages plus a system prompt, so memory and inference input are the same data.

Variant	Where Memory Lives	Persisted State Fields
`ai-agent`	JSON actor state	`messages`, `status`
`sandbox-coding-agent`	JSON actor state plus the sandbox ACP session	`messages`, `status`, `sessionId`
`experimental-durable-streams-ai-agent`	Durable streams; the actor stores only its conversation id and a read cursor	`conversationId`, `promptStreamOffset`

Message Handling

In the ai-agent example, the client pushes user input onto the agent's message queue with agent.connection.send("message", { text, sender }). This is a queue push, not an action call. The actor's run hook (see Lifecycle) consumes the queue serially with for await (const queued of c.queue.iter()).

Serial queue consumption is the per-conversation concurrency guarantee: at most one in-flight model call per actor, with no extra locking. The status field (thinking while a model call is in flight) is UI signal only; the run loop is the actual lock. The loop also checks c.aborted inside the token stream so shutdown exits gracefully.

Variant	Message Ingress	Serialization Guarantee
`ai-agent`	`message` queue pushed via `connection.send`	`run` hook pops one queued message at a time with `c.queue.iter()`.
`sandbox-coding-agent`	`sendMessage` action, no queue	Each call awaits the sandbox round trip before broadcasting the result.
`experimental-durable-streams-ai-agent`	Durable prompt stream long-polled from `onWake`	`promptStreamOffset` is persisted per chunk, so restarts resume without reprocessing prompts.

Streaming Responses

The ai-agent actor broadcasts a response event for every model text delta. The payload carries messageId, the per-token delta, the cumulative content, and a done flag (plus error on failure), so clients can either append deltas or idempotently replace the message by messageId using content. The example frontend replaces by messageId, which tolerates dropped events. The terminal broadcast has an empty delta, the full content, and done: true.

Because the assistant message object lives in c.state.messages and is mutated in place during streaming, partial content persists if the actor restarts mid-stream. The example broadcasts once per AI SDK delta with no throttling; batching or throttling deltas is a recommended extension for high-traffic deployments, not something the example implements.

Variant differences: sandbox-coding-agent sends a single response broadcast with done: true after the sandbox finishes (no incremental streaming), and experimental-durable-streams-ai-agent appends per-token chunks to a durable response stream, then broadcasts responseComplete or responseError.

Architecture

Topic	Summary
Topology	`agentManager["primary"]` singleton directory plus one `agent[agentId]` actor per conversation.
Ingress	Client pushes `AgentQueueMessage` payloads onto the agent's `message` queue with `connection.send`.
Streaming	One `response` broadcast per model delta, terminal broadcast with `done: true`.
Memory	Full transcript and status in JSON actor state; no external database.

The manager creates AgentInfo records and warms each agent through actor-to-actor communication: createAgent calls c.client<typeof registry>(), then client.agent.getOrCreate([info.id]) and awaits getStatus() so the conversation actor exists before the client connects. The sandbox variant extends this topology with a codingSandbox actor that shares the agent's key (codingSandbox.getOrCreate([c.key[0]])), so the agent-to-sandbox mapping is implicit in the key space.

Actors

Key: agentManager["primary"]
Responsibility: Directory actor. Creates AgentInfo records, lists agents, and warms each agent actor via c.client().
Actions
- createAgent
- listAgents
Queues
- None
State
- JSON
- agents
Key: agent[agentId]
Responsibility: One actor per conversation. Holds the full message history and status, consumes queued user messages in its run loop, calls the model via the AI SDK, and broadcasts streaming deltas.
Actions
- getHistory
- getStatus
Queues
- message
Events
- messageAdded
- status
- response
State
- JSON
- messages
- status

Lifecycle

sequenceDiagram
	participant C as Client
	participant AM as agentManager
	participant A as agent
	participant LLM as Model API

	C->>AM: createAgent(name)
	AM->>A: getOrCreate([info.id]) + getStatus()
	AM-->>C: AgentInfo
	C->>A: connection.send("message", {text, sender})
	Note over A: run loop pops queue via c.queue.iter()
	A-->>C: messageAdded (user message)
	A-->>C: messageAdded (assistant placeholder)
	A-->>C: status (thinking)
	A->>LLM: streamText(system prompt + history)
	loop each text delta
		LLM-->>A: delta
		A-->>C: response {messageId, delta, content, done: false}
	end
	A-->>C: response {delta: "", content, done: true}
	A-->>C: status (idle)

Security Checklist

The examples ship without auth so they stay minimal. Apply this baseline before exposing an agent backend.

API keys stay server-side: OPENAI_API_KEY (or ANTHROPIC_API_KEY) is read by the AI SDK inside the actor process. The key never reaches the browser; clients only talk to the actor over RivetKit. The sandbox variant forwards keys into the sandbox env, never to the client.
Add authentication: The examples have no auth, so anyone who reaches the server can create agents, list them, and message any agent whose key they can guess. Add onBeforeConnect or createConnState checks with scoped tokens as a recommended extension. See Authentication.
Validate and rate-limit queue payloads: The example only skips bodies without a string text. Enforce payload size limits, schema validation, and per-connection rate limits as a recommended extension.
Derive sender identity server-side: The example trusts the client-supplied sender field verbatim. Bind sender identity to the authenticated connection instead.
Cap or trim message history: The example sends the full transcript on every model call with no cap. Trim or summarize old messages as a recommended extension so prompts and state stay bounded.
Set cost ceilings per conversation: Add per-agent token budgets and quotas as a recommended extension. The sandbox variant runs real compute, so also enforce per-user sandbox quotas and restrict sandbox network egress.