README.md

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
Support Resolution	Non-JSON pipeline	Classify, route, and draft responses using output_field — works with any model including local Ollama

Providers

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

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 builds and serves 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
uv sync
pytest

See CONTRIBUTING.md. Report bugs via Issues.

License

Apache License 2.0