Skip to content

Human-in-the-loop

Put a human inside an agent pipeline — as an approval gate, a form, or an interactive REPL — without leaving the Agent abstraction. LazyBridge ships two human engines and two sugar factories that return a plain Agent:

  • Human approval — HumanEngine / human_agent. A single-turn step: present a task to a person, collect their response (free text or a Pydantic form), return it as an Envelope. Terminal or web UI.
  • Supervisor REPL — SupervisorEngine / supervisor_agent. An interactive REPL the human drives: approve/continue, retry a named sub-agent with feedback, call tools, read the store — with an audit trail.

Ships in lazybridge.ext.hil. Both are Engine classes you pass via Agent(engine=…) (like LLMEngine or Plan); the human_agent(...) / supervisor_agent(...) factories are the one-liner sugar and return an Agent. HiL lives in ext rather than core because it's a workflow pattern, not a primitive every agent needs. Import from lazybridge.ext.hil.

When to use which

You need… Use
A single approval / form / clarification turn HumanEngine / human_agent
An interactive session: tools, agent-retry, store access, audit trail SupervisorEngine / supervisor_agent (the REPL)
Only to check an LLM's output (no human) verify= on a normal agent — not HiL
Unattended automation with a human fallback Either engine with timeout= + default= set

They behave as agents

human_agent(...) and supervisor_agent(...) return a real Agent — callable, composable, and usable as a step or a tool:

from lazybridge import Agent, Plan, Step
from lazybridge.ext.hil import human_agent

ask_city = human_agent(name="ask_city")          # terminal UI by default
pipeline = Agent(engine=Plan(Step(gather),
                             Step(ask_city, task="What city is the user in?"),
                             Step(finalise)),
                 name="address_normaliser")

Everything you already know — chain, parallel, Plan, routing — still works around a human engine; it's just another engine.

See also