Python AI Agent Workflow Templates — Pro

WHAT THIS IS

Small, readable Python modules for the LLM workflows most projects end up writing by hand. One thin client layer talks to any OpenAI-compatible endpoint (OpenAI, Groq, Together, Ollama, vLLM, LM Studio) or to Anthropic, so the templates don't care which provider you use. There's no LangChain or other agent framework, just pydantic plus the official SDK you choose.

Tested on Python 3.13 with the openai 3.41 and anthropic 2.3 SDKs. Needs Python 3.11 or newer.

ALSO INCLUDES EVERYTHING IN STANDARD ($29)

- Tool-calling agent loop: register Python functions as tools, max-step guard, tool errors returned to the model

- Structured extraction with Pydantic: schema in the prompt, validation, bounded repair retries

- RAG-lite: chunk local .md/.txt/.py files, BM25 ranking with no vector DB, answers cite their sources

- Summarize-and-route pipeline: summary, then a classified route with a confidence threshold and fallback, then your handler

- Retries and cost logging: exponential backoff with jitter, retries on 429 and 5xx, token and cost ledger, JSONL log, budget cap

- FakeLLM for offline tests (scripted replies, simulated failures, call recording)

- OpenAI-compatible and Anthropic adapters, tested against the SDKs' real response types

- 19 pytest tests, ruff clean, and a runnable offline demo

- License: 1 developer

WHAT PRO ADDS ($59)

- Eval harness: cases and checks (contains, equals, regex, attribute, LLM-as-judge), pass rate, JSON report for CI

- FastAPI service wrapper: extract, route, ask and usage endpoints, public /healthz, bearer-token auth

- Concurrency with rate limiting: concurrency cap, token bucket for requests/min and tokens/min, results kept in order, per-item errors

- Human-approval step: pauses before risky tools such as refunds or emails, saves state to JSON, approve or reject, then resume

- Dockerfile (slim, non-root, healthcheck)

- CI workflow for GitHub: ruff and pytest on Python 3.11 to 3.13, then a Docker build and auth smoke test

- 30 pytest tests

- License: 1 developer or 1 team in one organization

HONEST NOTES

- These are templates, not a framework. You're expected to read and edit them.

- All tests run offline with a fake model. The provider adapters are tested against the SDKs' real response types, but no live API calls were made during testing, so try your own model before shipping.

- Prices in the cost logger are placeholders you fill in. Nothing is a quoted provider price, and model names in the README are only examples.

- RAG-lite uses keyword (BM25) search. It's fast and has no dependencies, but it isn't semantic search. Swap in embeddings if you need that.

- The OpenAI adapter uses Chat Completions, the endpoint OpenAI-compatible servers implement, not the newer Responses API.

- Pro's Dockerfile and CI workflow follow the standard formats, but the Docker image wasn't built during testing. The service itself was tested, including a live uvicorn run.

FAQ

Do I need an API key to try it? No. The tests, the demo and the service's default mode use a fake model.

Can I use it for client work? Yes, within your license (Standard: 1 developer; Pro: 1 developer or 1 team).

Refunds? Yes, 14-day refund policy: if the templates aren't useful to you, email within 14 days of purchase for a full refund, no questions asked.

Updates? Re-download from your Payhip library when a new version is posted.

Support: goodspress@mail.grokbot.com