Cookbook

Every concept in these docs has a runnable example in the CARL repo. This is the index; each links to a complete, executable script.

Running examples

# Set an API key for examples that call an LLM
export OPENAI_API_KEY="sk-or-v1-..."

# Run one example
PYTHONPATH=$(pwd) uv run python examples/orchestration/basic_chain_example.py

# Or run them all
make examples

Tool-only and metric examples run without an API key.

Orchestration

Example	Shows
basic_chain	Core chain, dependencies, serialization.
parallel_branches	DAG parallelism across independent steps.
conditions	Conditional branching.
loop_until	Loop-back until a condition holds.
execution_modes_mock / pipeline	FAST vs SELF_CRITIC.

Tool calling

Example	Shows
tool_steps	Tool steps, memory, mixed chains.
structured_output	Schema-constrained JSON output.

Agents (multi-agent)

Example	Shows
supervisor_routing	LLM routing to specialists.
llm_council	N-sample voting.
human_in_the_loop	Pause for human input.
agent_skill	AgentSkill execution.

Evaluation

Example	Shows
metrics	Custom + built-in metrics (no API key).
dataset_evaluator	Batch evaluation + report.
reflection / reflection_metrics	Reflection, metric-fed reflection.

RE-PLAN

Example	Shows
deterministic · llm_checker · voting · checkpoint_rollback · budget_guard	The RE-PLAN strategies.

Skills & LLM inference

Example	Shows
chain_from_description	Generate a chain from NL.
skill_resolver	URI-based skill resolution.
openrouter · token_usage	OpenRouter, token accounting.

Build one yourself

See the end-to-end tutorial — a multi-step agent built from scratch, combining LLM, tool, memory, and conditional steps.