Fixed-order calls#

Call the LLM (optionally) while invoking multiple sub-functions in an order hard-coded in Python.

When to use#

• Research pipeline: survey → find gaps → generate ideas
• Paper pipeline: draft → review → revise
• Data pipeline: collect → clean → analyze
• Any multi-step task whose step order is known ahead of time

Design points#

• Use the @agentic_function decorator
• Call multiple sub-@agentic_functions in a fixed order
• exec() is optional: skip it (pure chaining), or call it multiple times (each call creates one exec child node)
• Data flows between sub-functions through plain Python variables
• One function may call exec() multiple times AND call any number of other @agentic_functions

Example: no exec, pure chaining#

from openprogram import agentic_function

@agentic_function(input={
    "task": {"description": "Research topic."},
})
def research_pipeline(task: str, runtime=None) -> dict:
    """Run the full research pipeline: survey, find gaps, generate ideas."""
    survey = survey_topic(topic=task, runtime=runtime)
    gaps = identify_gaps(survey=survey, runtime=runtime)
    ideas = generate_ideas(gaps=gaps, runtime=runtime)

    return {"survey": survey, "gaps": gaps, "ideas": ideas}

Example: one exec call to summarise#

@agentic_function(input={
    "task": {"description": "Research topic."},
})
def research_pipeline(task: str, runtime=None) -> str:
    """Run the full research pipeline and summarise the results."""
    survey = survey_topic(topic=task, runtime=runtime)
    gaps = identify_gaps(survey=survey, runtime=runtime)
    ideas = generate_ideas(gaps=gaps, runtime=runtime)

    return runtime.exec(content=[
        {"type": "text", "text": (
            f"Survey:\n{survey}\n\n"
            f"Gaps:\n{gaps}\n\n"
            f"Ideas:\n{ideas}"
        )},
    ])

Session DAG#

Each call is one node; the caller edge points at the orchestrator:

research_pipeline
├── survey_topic       ← step 1
├── identify_gaps      ← step 2
└── generate_ideas     ← step 3

Passing data between steps#

Sub-functions hand data to each other through Python variables — no LLM involved:

survey = survey_topic(topic=task, runtime=runtime)
gaps = identify_gaps(survey=survey, runtime=runtime)

The return value of survey_topic goes straight in as the input argument of identify_gaps.

Inserting Python processing between steps#

survey = survey_topic(topic=task, runtime=runtime)

# plain Python processing in between
key_points = extract_key_points(survey)
filtered = [p for p in key_points if p["relevance"] > 0.5]

gaps = identify_gaps(survey="\n".join(filtered), runtime=runtime)

Error handling#

The primary mechanism is exception propagation: when a sub-function raises, its DAG node is recorded with status='error' and the exception re-raises into the orchestrator. Catch it there with a plain try/except:

try:
    survey = survey_topic(topic=task, runtime=runtime)
except Exception as e:
    return {"error": f"Survey failed: {e}"}

gaps = identify_gaps(survey=survey, runtime=runtime)

Optionally, if a sub-function reports failure in-band (returning an error string instead of raising), check the value:

survey = survey_topic(topic=task, runtime=runtime)
if not survey or "error" in survey.lower():
    return {"error": "Survey failed", "survey": survey}

gaps = identify_gaps(survey=survey, runtime=runtime)

Versus "LLM-selected calls"#