02 Chaining output to input

Before you start

Complete 01 Linear pipelines. Read 04 Context and prior steps if you have not already. That lesson introduces ContextPolicy; here we focus on why handoffs matter in multi-step workflows.

Concept

In a linear pipeline, each step produces text output. When the runtime builds the prompt for a later step, it can include:

This is how step two "sees" step one's work without you manually copying strings in Python. The runtime assembles context before the agent runs.

Control that assembly with ContextPolicy on each Agent:

Common pattern for research then write:

• Researcher: include_prior_steps="none", include_run_input=True (focus on the user topic).
• Writer: include_prior_steps="last", include_run_input=False (summarize research only, not the raw topic).

If you omit context, the runtime applies its default policy. Explicit policies make handoffs predictable before you switch from stub to a paid provider.

You do not pass prior output manually in run(). The chain is declarative: register steps in order and set policies per agent.

Example

Save as chain_handoff.py:

from arcflow import Agent, ContextPolicy, Workflow

researcher = Agent(
 name="researcher",
 role="Research",
 instructions="Research the topic and list key points.",
 context=ContextPolicy(
 include_prior_steps="none",
 include_run_input=True,
 ),
)

writer = Agent(
 name="writer",
 role="Writer",
 instructions="Write a summary based only on the research output.",
 context=ContextPolicy(
 include_prior_steps="last",
 include_run_input=False,
 max_prior_step_chars=8192,
 ),
)

workflow = Workflow("chain-handoff")
workflow.step(researcher).step(writer)

result = workflow.run("Microgrids for remote clinics")

print(result.output[:200], "..." if len(result.output) > 200 else "")
print(f"status={result.status} steps={result.step_count}")

Run:

python chain_handoff.py

The researcher consumes the run input. The writer is instructed to rely on prior step text; with include_run_input=False, the writer should not re-read the original topic verbatim in its prompt assembly.

Verify

Sanity check for invalid policy:

from arcflow import ContextPolicy
from arcflow.exceptions import WorkflowConfigurationError

try:
 ContextPolicy(include_prior_steps="sometimes")
except WorkflowConfigurationError:
 print("Invalid include_prior_steps rejected as expected")

Next

03 Graph workflows intro covers conditional routing when linear order is not enough.