pi-agent

Agent runtime with tool calling. Part of the pi agent harness — a Rust port of @earendil-works/pi-agent-core.

Builds on pi-ai and adds:

• run_agent / run_agent_with_history — streaming agent loop with sequential tool execution.
• AgentTool trait — implement once, plug in via AgentConfig::with_tools.
• PermissionPolicy trait with Allow / AllowSession / Deny — tools that flag requires_permission() (e.g. bash, write, edit) are gated.
• Streaming AgentEvent::TextDelta / ThinkingDelta / ToolExecutionStart / ToolExecutionEnd events through an mpsc::UnboundedSender.
• Built-in tools under pi_agent::tools::*: read, write, edit, bash, ls, grep, glob_tool, web_fetch, todo. Pre-bundled as pi_agent::tools::default_tools().

Quick start

use pi_agent::{run_agent, tools::default_tools, AgentConfig};
use pi_ai::{Message, Model};

# tokio_test::block_on(async {
let cfg = AgentConfig::new(
    Model::anthropic_claude_sonnet_4_6(),
    "You are a helpful coding assistant.",
)
.with_tools(default_tools())
.with_max_turns(8);

let user = Message::user_text("List the .rs files in the current directory.");
let run = run_agent(&cfg, user, None).await.unwrap();
println!("{} messages", run.messages.len());
# });

License

MIT