Your First Flow¶

Run a single agent, then a parallel fanout, then a DAG flow — five commands total.

Single agent turn¶

li agent claude "What are the three laws of thermodynamics?"

The model's response prints to stdout. Continue the same conversation:

li agent -c "And what is the zeroth law?"

Parallel fanout¶

li o fanout claude "audit the file structure in this directory" -n 3 --save ./results

3 workers run in parallel. Artifacts land in ./results/ — one file per worker.

DAG flow¶

Preview the plan before executing:

li o flow claude "research Python async best practices and write a summary" --dry-run

# output:
Planning DAG...
FlowPlan (2 agents, 3 ops, synthesis=True)

Agents:
  r1: researcher
  c1: critic

Operations:
  o1 → r1
  o2 → r1    depends_on: o1
  o3 → c1    depends_on: o2  [CONTROL]

Model resolution:
  r1: codex/gpt-5.4 (profile)
  c1: claude_code/opus-4-7 (profile)

Execute:

li o flow claude "research Python async best practices and write a summary" --save ./results

Run directory¶

Every li invocation writes to ~/.lionagi/runs/{run_id}/:

~/.lionagi/runs/
└── 20260420T103404-a1b2c3/
    ├── run.json       # manifest: command, branches, agents
    ├── branches/      # branch snapshots
    └── artifacts/     # agent output files

run_id format: YYYYMMDDTHHMMSS-{uuid6}. Resume any branch: li agent -r <branch_id> "...".

Python API¶

import asyncio
from lionagi import Branch

async def main():
    b = Branch()
    response = await b.operate(instruction="What is the observer pattern?")
    print(response)

asyncio.run(main())

Branch.operate() returns the model response as a string. The CLI and Python share the same internals.

Next: CLI reference for full flag tables, or Cookbook scenarios.