04 Context and prior steps

Before you start

Complete 03 Roles and multi-agent pipelines. You should have a two-step pipeline running with the stub provider and understand that step order determines execution sequence.

Concept

When the runtime builds the prompt for an agent, it can include:

These three fields form a ContextPolicy. Pass it to Agent(..., context=ContextPolicy(...)).

include_prior_steps

include_run_input

When True, the agent still sees the original run input. When False, later steps rely on prior step output only. A writer that should summarize research but ignore the raw user message often sets include_run_input=False.

max_prior_step_chars

Caps how much prior text enters the prompt (minimum 256 in the Python SDK). Prevents runaway prompt size when intermediate steps produce long output.

Default behavior when you omit context is defined by the runtime. Explicit policies make multi-agent handoffs predictable before you connect a paid provider.

Example

Researcher sees the user topic but no prior steps (it runs first). Writer sees the researcher's output but not the raw run input.

Save as context_policy_demo.py:

from arcflow import Agent, ContextPolicy, Workflow

researcher = Agent(
 name="researcher",
 role="Research",
 instructions="Research the given topic thoroughly.",
 context=ContextPolicy(
 include_prior_steps="none",
 include_run_input=True,
 max_prior_step_chars=4096,
 ),
)

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

workflow = Workflow("context-demo")
workflow.step(researcher)
workflow.step(writer)

result = workflow.run("Analyze renewable energy trends")

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

Run:

python context_policy_demo.py

The researcher policy keeps the first step focused on the user input without empty prior-step noise. The writer policy forces it to work from the researcher's output rather than re-reading the original topic verbatim.

Verify

Sanity-check invalid policy values:

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")

Trace metadata records sizes, not prompt content. After a run, ProviderRequestSent (when using a real provider) includes prompt_size_bytes; no instruction or user text is logged (trace data policy).

Next