Marketing Content Agent Example

A marketing content agent whose prompts get better over time. Built on the full Autoresearch Prompt Manager stack, powered by agno.

Install

pip install autoresearch-prompt-manager[example]

Configure

export PM_LLM_PROVIDER=groq              # or: anthropic, openai, gemini, openrouter
export PM_LLM_MODEL=openai/gpt-oss-120b  # or: claude-sonnet-4-20250514, gpt-4o, etc.
export PM_LLM_API_KEY=your-api-key       # required
export PM_DATABASE_URL=postgresql://prompt_manager:prompt_manager@localhost:15432/prompt_manager

Quick start

# Start the API
arpm-api up       # Start PostgreSQL
arpm-api start    # Run migrations + start API on :8910

# Seed prompt templates
arpm-example seed

# Generate content
arpm-example run "Write a welcome email for Alice joining TechCorp"

# Run the full autoresearch optimization loop
arpm-example loop

# Check status
arpm-example status

Commands

What it demonstrates

1. Prompt resolution -- the agent calls arpm-example run and gets the best template via experiment-aware routing
2. A/B experiment routing -- MurmurHash3 deterministic routing, same user always gets the same variant
3. Quality metrics -- the agent self-evaluates and reports scores back to the API
4. Autonomous optimization -- autoresearcher-shonku analyses metrics, proposes improved versions, deploys experiments with adjusted weights

The full loop (arpm-example loop)

This runs 6 steps:

Step 1. Create a prompt with 2 versions (formal vs casual)

Step 2. Create an A/B experiment with 50/50 routing

Step 3. Run the marketing agent 4 times. Each session gets routed to a different version. The agent generates content, rates it, reports the metric.

Step 4. Check metrics per version:

v1 (formal): mean=6.00
v2 (casual): mean=6.50

Step 5. Autoresearcher runs. It reads the metrics, reads the prompts, proposes v3 (combining the best of both), validates safety, and deploys a new experiment:

Step 6. Verify all 3 versions are now receiving traffic.

How tools flow

arpm-example (this package)
  │
  │ defines tools that wrap the prompt-manager API:
  │   resolve_prompt  → GET /resolve/{slug}
  │   report_metric   → POST /metrics
  │
  └─→ MarketingContentAgent (shonku agent)
        │
        │ agent's built-in tool:
        │   rate_content  → heuristic quality scorer
        │
        └─→ agno → LLM (Groq gpt-oss-120b)

For the optimization loop, the autoresearcher gets 6 additional tools (get_prompt, get_metrics, create_version, create_experiment, conclude_experiment, get_sample_interactions) that also wrap the API.

The autoresearcher has no built-in knowledge of your prompts. It learns everything through tool calls.

Seed prompts

Contributing

# Clone the repo
git clone git@github.com:kaustav1996/autoresearch-prompt-manager.git
cd autoresearch-prompt-manager/packages/example
pip install -e ".[dev]"
pytest

License

MIT