agent-actions 0.1.1

Declarative YAML-based framework for orchestrating LLM workflows with batch processing

Declarative LLM orchestration. Define workflows in YAML — each action gets its own model, context window, schema, and pre-check gate. The framework handles DAG resolution, parallel execution, batch processing, and output validation.

[!WARNING] Experimental — Under active development. Expect breaking changes. Open an issue with feedback.

actions:
  - name: extract_features
    intent: "Extract key product features from listing"
    model_vendor: anthropic              # Each action picks its own model
    model_name: claude-sonnet-4-20250514

  - name: generate_description
    dependencies: [extract_features]
    model_vendor: openai                 # Mix vendors in one pipeline
    model_name: gpt-4o-mini
    context_scope:
      observe:
        - extract_features.features      # See only what it needs
      drop:
        - source.raw_html                # Don't waste tokens on noise

Install

pip install agent-actions

Quick start

agac init my-project && cd my-project                # scaffold a project
agac init --example contract_reviewer my-project     # or start from an example
agac run -a my_workflow                              # execute

Why not just write Python?

You will, until you have 15 steps, 3 models, batch retry, and a teammate asks what your pipeline does.

Capability	Agent Actions	Python script	n8n / Make
Per-step model selection	YAML field	Manual wiring	Per-node config
Context isolation per step	observe / drop	You build it	Not available
Pre-check guards (skip before LLM call)	guard:	If-statements	Post-hoc branching
Parallel consensus (3 voters + merge)	2 lines of YAML	Custom code	Many nodes + JS
Schema validation + auto-reprompt	Built in	DIY	Not available
Batch processing (1000s of records)	Built in	For-loops	Loop nodes
The YAML is the documentation	Yes	No	Visual graph

Examples

Example	Pattern	Key Features
Review Analyzer	Parallel consensus	3 independent scorers, vote aggregation, guard on quality threshold
Contract Reviewer	Map-reduce	Split clauses, analyze each, aggregate risk summary
Product Listing Enrichment	Tool + LLM hybrid	LLM generates copy, tool fetches pricing, LLM optimizes
Book Catalog Enrichment	Multi-step enrichment	BISAC classification, marketing copy, SEO metadata, reading level
Incident Triage	Parallel consensus	Severity classification, impact assessment, team assignment, response plan

Providers

Provider	Batch	Provider	Batch
OpenAI	Yes	Groq	Yes
Anthropic	Yes	Mistral	Yes
Google Gemini	Yes	Cohere	Online only
Ollama (local)	Online only

Switch providers per-action by changing model_vendor.

Key capabilities

• Pre-flight validation — schemas, dependencies, templates, and credentials checked before any LLM call
• Batch processing — route thousands of records through provider batch APIs
• User-defined functions — Python tools for pre/post-processing and custom logic
• Reprompting — auto-retry when LLM output doesn't match schema
• Observability — per-action timing, token counts, and structured event logs
• Interactive docs — agac docs serve generates a visual workflow dashboard

Documentation

• Full docs — guides, tutorials, reference
• Configuration — YAML schema reference
• CLI — all commands and options

Contributing

git clone https://github.com/Muizzkolapo/agent-actions.git && cd agent-actions
pip install -e ".[dev]"
pytest

See CONTRIBUTING.md. Report bugs via Issues.

License

Apache License 2.0