continuum-workflows

Continuum Workflow Agents Skill

Workflow agents are themselves BaseAgent subclasses, so they nest. Each ships with a create_* factory.

Authoritative source: docs/agent.md, §6.

---

Imports

from continuum.agent import (
    create_sequential_agent, create_parallel_agent, create_loop_agent,
    create_reflection_agent, create_planner_agent, create_router_agent,
    MemoryScope, MergeStrategy, FailStrategy, TerminationType,
)
# Debate, Scatter, and Supervised factories live one level deeper —
# they are NOT re-exported from `continuum.agent`.
from continuum.agent.workflow import (
    create_debate_agent, create_scatter_agent, create_supervised_agent,
)

---

SequentialAgent

pipeline = create_sequential_agent(
    name="content-pipeline",
    agents=[researcher, writer, editor],
    pass_full_history=False,
    fail_strategy=FailStrategy.FAIL_FAST,
)

Use when each agent's output should feed the next.

---

ParallelAgent

fanout = create_parallel_agent(
    name="fanout",
    agents=[a, b, c],
    merge_strategy=MergeStrategy.LLM_SUMMARIZE,    # CONCATENATE / STRUCTURED / FIRST_SUCCESS
    fail_strategy=FailStrategy.CONTINUE_ON_ERROR,
    timeout=300,
)

Use when independent perspectives can run concurrently and you want them merged.

---

LoopAgent

iterate = create_loop_agent(
    name="iterate-until-done",
    agent=worker,
    termination_type=TerminationType.LLM_DECISION,
    max_iterations=10,
    termination_prompt="Reply COMPLETE if done, CONTINUE otherwise.",
    # or termination_tool="finish" / termination_pattern="^DONE" /
    #     termination_condition=fn(content, history)
)

---

ReflectionAgent

self_improving = create_reflection_agent(
    name="self-improving",
    agent=writer,
    critique_prompt=None,                  # default critique
    max_reflections=2,
    reflection_model=None,                 # defaults to writer's model
)