Workflow Patterns

AutoFlow's pattern matcher (src/autoflow/patterns/matcher.py) analyzes normalized task contracts and recommends the best workflow pattern. The GeneratorAgent includes this recommendation as a RECOMMENDED PATTERN: hint in the generation prompt.

Pattern Reference

Overview

The following patterns are used by AutoFlow to select an appropriate workflow structure based on task intent, keywords, and normalized task contracts.

simple_qa

Use case: Quick question-and-answer tasks with a single LLM call.

Typical structure: 1–2 nodes (intake LLM → optional formatter)

Keywords: question, answer, explain, what is, how to

extract_format

Use case: Reading files or documents and extracting structured data.

Typical structure: LLM extraction (with file_variables: input.source_file) → variable_transform

Keywords: extract, parse, read, document, file, PDF, CSV

multi_step

Use case: Sequential processing pipelines where each step builds on the previous.

Typical structure: LLM₁ → LLM₂ → LLM₃ → response (chain of 3+ nodes)

Keywords: step, pipeline, sequence, then, first, next, finally

batch_processing

Use case: Processing multiple items from a list using iteration.

Typical structure: input via file_variables → loop → process item → collect results

Keywords: batch, list, each, every, multiple, iterate, bulk, all

conditional

Use case: Binary decision routing based on a condition.

Typical structure: classify → if_else → branch A / branch B

Keywords: if, condition, check, verify, when, based on, depends

classification

Use case: Multi-way routing based on category classification.

Typical structure: intake → decision_tree → category-specific handlers

Keywords: classify, categorize, sort, route, triage, priority, type

transform_chain

Use case: Data transformation and reformatting pipelines.

Typical structure: input → variable_transform → LLM format → output

Keywords: transform, convert, format, reformat, map, change, reshape

self_improving

Use case: Evaluation and iterative improvement workflows.

Typical structure: generate → evaluate → if_else (pass/fail) → critique → regenerate

Keywords: evaluate, review, improve, iterate, quality, score, assess, refine

How Pattern Matching Works

Overview

Pattern matching is applied after task normalization and before workflow generation.

1. The NormalizerAgent produces a structured task contract
2. The pattern matcher evaluates keyword rules in priority order against the task contract and user input
3. The best-matching pattern (with confidence score) is selected
4. The GeneratorAgent receives a RECOMMENDED PATTERN: hint and a MODEL RECOMMENDATION: hint based on the pattern-to-task-type mapping

Pattern → Task Type Mapping

Recommended models come from the canonical Ask Sage model catalog (node_registry/model_catalog.yaml), which maps each pattern to a role (e.g. default, cheap_extractor, analyzer, classifier, deliverable_writer, evaluator, creative) and each role to a concrete model. Concrete models include gpt-5.4, google-claude-46-sonnet, google-claude-45-opus, gpt-4o, gpt-4o-mini, and auto.

These model recommendations appear as hints in the generated workflow's LLM node configs, not as the model running the pipeline.

Use python -m autoflow --list-models to print the current catalog, --lint-models to find templates using uncatalogued names, and --remap-model OLD=NEW to bulk-rename across templates.

Golden Workflows

AutoFlow includes 8 hand-crafted golden reference workflows in workflows/golden/ that demonstrate correct v1 (real Ask Sage) format. Each has fully wired handles, terminal nodes, resolvable variable references, and an agent_config block ready for Agent Builder import.

Golden workflows pass full validation (schema + semantic) and registry config checks. They serve as reference implementations for each pattern.