autumn-plugin-agent

Provider-agnostic LLM agent harness with tool calling for Autumn apps.

Requires autumn-web 0.10.

One trait, two providers, zero heavyweight LLM SDKs. The plugin speaks HTTP directly:

• OpenAI-compatible chat-completions (POST /chat/completions) — works with OpenAI, Ollama, vLLM, LM Studio, and any proxy that exposes the standard OpenAI protocol.
• Anthropic Messages API (POST /v1/messages).

Quickstart

Add the plugin and a tool:

use autumn_plugin_agent::plugin::{AgentHandle, AgentPlugin};
use autumn_plugin_agent::{AgentError, FnTool};
use autumn_web::prelude::*;
use std::sync::Arc;

async fn get_weather(input: serde_json::Value) -> Result<serde_json::Value, AgentError> {
    let city = input["city"].as_str().unwrap_or("unknown");
    Ok(serde_json::json!({"city": city, "temp_f": 72}))
}

#[get("/ask")]
async fn ask(agent: AgentHandle) -> String {
    agent
        .run("What is the weather in Chicago?")
        .await
        .map(|o| format!("{o:?}"))
        .unwrap_or_else(|e| e.to_string())
}

#[autumn_web::main]
async fn main() {
    let weather = FnTool::new(
        "get_weather",
        "Current weather for a city.",
        serde_json::json!({
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        }),
        get_weather,
    )
    .shared();

    autumn_web::app()
        .plugin(AgentPlugin::new().tool(weather))
        .routes(routes![ask])
        .run()
        .await;
}

Configure in autumn.toml:

[agent]
provider = "openai-compatible" # or "anthropic"
model = "gpt-4o-mini"
# base_url = "http://localhost:11434/v1"  # Ollama, vLLM, ...
max_steps = 10
max_tokens = 32000
# max_run_secs = 300        # wall-clock limit per run
# prompt_caching = true     # Anthropic cache breakpoints (default on)

The API key comes only from the environment — never from the config file:

export AGENT_API_KEY="sk-..."

Always-on agents

The crate has the primitives that today's always-on agents (Hermes, OpenClaw, Dots, Muse) share. See docs/research/always-on-agents.md.

use std::sync::Arc;
use std::time::Duration;
use autumn_plugin_agent::{AgentPlugin, Heartbeat, InMemoryMemoryStore, ToolRules};

AgentPlugin::new()
    .tool(send_email)                                  // declares ToolEffect::External
    .memory_store(InMemoryMemoryStore::default().shared())
    .policy(Arc::new(ToolRules::ask_before_acting()))  // pause before writes
    .heartbeat(Heartbeat::every(Duration::from_secs(1_800)).session("ops"))
    .followups(Duration::from_secs(24 * 3_600))        // agent books its own wake-ups
    .delivery(my_slack_delivery);                      // where reports go

Run cargo run --example always_on_agent to see memory, an approval round trip, and two heartbeat ticks against a scripted model.

What the plugin installs

• [agent] config section (declared strict-config-safe).
• agent_run and agent_resume background jobs — enqueue with enqueue_agent_run(AgentRunArgs::new(prompt)), enqueue_agent_run_tracked, or enqueue_agent_resume (re-exported at the crate root).
• agent_heartbeat scheduled task, when a Heartbeat is set.
• agent health indicator on /actuator/health — a cheap list_models ping against the same provider client the running app uses (shared with the startup hook; DOWN until startup completes). It sits in the health-only group, so a sick provider never blocks deploys.
• AgentHandle extractor for handlers — a ready-to-run agent from app state.

The agent loop

Agent::run iterates: call model → execute tool calls → append results → repeat, until the model answers, a budget runs out, the loop guard fires, or the policy pauses for approval. max_steps counts tool-execution rounds (the final answer is free) and max_tokens covers estimated plus provider-reported usage; the per-call output cap reserves the estimated history cost first, and a final answer that overshoots the token budget still returns AgentOutcome::BudgetExhausted, not an error. Unknown tools, tool failures, denials, and rejections reach the model as error payloads so it can recover. Calls in one round run concurrently (parallel_tool_calls(false) runs them one by one). When the history must shrink, the loop keeps the system prompt and the task message and never leaves a tool result without its call.

Background runs

use autumn_plugin_agent::{AgentRunArgs, SessionId, enqueue_agent_run_tracked};

let handle = enqueue_agent_run_tracked(
    AgentRunArgs::new("Summarize today's signups.")
        .session(SessionId::new("daily-report"))
        .deliver(true),
)
.await?;
// Poll handle.status_path(); the result is the AgentOutcome as JSON.

The run id is fixed at enqueue time, so a retried job keeps it. Tools read it from ToolContext to build idempotency keys.

Crate layout

Known issues

• The default session store and the in-memory memory store lose data on restart. Supply database-backed SessionStore / MemoryStore implementations in production.
• A retried agent_run job re-runs the whole loop. Step checkpoints are a follow-up; tools can deduplicate on ToolContext::run_id meanwhile.
• Non-streaming only. Streaming chat completions is an explicit follow-up (see CHANGELOG).
• Token accounting mixes provider-reported usage with a bytes / 4 heuristic between calls. Good enough for a guardrail, never for billing.
• The plugin reads [agent] from autumn.toml itself with the toml crate; Autumn 0.8 still exposes no per-section accessor to plugins, so the framework's own profile layering does not apply to plugin keys yet.

License

Apache-2.0.