oharness-loop

Agent, Loop trait, and shipped loop implementations for open-harness — the user-facing orchestration layer.

What's in here

• Agent + AgentBuilder — the top-level "I want to run an agent" entry point. Threads LLM, tools, memory, event sink, budget, critics, and a chosen Loop into a run.
• Loop trait — a loop owns the turn-taking policy. Shipped:
  • ReactLoop (feature react, default-on) — the classic reason-act-observe loop with tool dispatch.
  • ConversationLoop (feature conversation) — alternates assistant turns with a UserSimulator.
• run_reflexion (feature reflexion) — multi-episode wrapper that threads Reflector-produced notes into the next episode via ReflectionInjector.
• UserSimulator trait + ScriptedUserSimulator + LlmUserSimulator — the user-side model for conversational loops.

Quickstart — first agent in 10 lines

use oharness_loop::{Agent, ReactLoop};
use oharness_tools::fs::FsToolSet;
use oharness_core::Task;
use std::sync::Arc;

let agent = Agent::builder()
    .with_llm(my_llm)                          // impl Llm
    .with_tools(Arc::new(FsToolSet::new()))
    .with_loop(Box::new(ReactLoop::new()))
    .with_max_turns(10)
    .build()?;

let outcome = agent.run(Task::new("say hello")).await?;

See examples/hello_scripted.rs for a runnable version. The examples/ directory ships 11 runnable examples covering tools, critics, budgets, replay, reflexion, custom middleware, custom memory policies, and multi-agent conversation.

Feature flags

• react (default) — ReactLoop.
• conversation — ConversationLoop.
• reflexion — run_reflexion helper.
• llm-judge — forwards to oharness-critic/llm-judge so the llm_judge_critic example builds.

License

Dual-licensed under MIT or Apache-2.0.