Hi HN - I'm sharing Pixelagent (https://github.com/pixeltable/pixelagent), an open-source agent engineering blueprint we built to tackle the core infrastructure challenges of agent SDK development.

I see agents as the intersection of an LLM, storage, and orchestration. With this reference implementation, I show how you can unify them into a single declarative agentic framework.

Many frameworks focus on orchestration; for me the real pain points are in data infrastructure. This is why you see so many abstractions around memory for instance and people re-inventing the wheel around state management. All of these are just data storage, indexing, retrieval, and orchestration.

With Pixelagent, I show how you can get:

- Multimodal data handling: Processing text, images, audio, and video through a consistent interface

- Persistent state management and built-in observability: Automatic tracking and versioning of agent memory, tool calls, and context...

- Provider agnostic: Support for OpenAI, Anthropic, and AWS Bedrock with the same API.

- Declarative Model: A type-safe framework

- Agentic Extensions: Add reasoning, reflection, memory, knowledge, and team workflows.

Fork it, make your own or whatnot! I hope it inspires people to build the abstractions they need.

Onward,

Pierre

---

Get Started:

- Pixelagent GitHub Tutorial: https://github.com/pixeltable/pixelagent/tree/main/blueprint...

- Pixelagent Documentation: https://docs.pixeltable.com/docs/libraries/pixelagent

- Join our Discord: https://discord.gg/QPyqFYx2UN

Pixelagent: Agent Engineering Blueprint https://www.youtube.com/watch?v=BS6PRsnxkBA

Exciting to see a fellow builder in the space! Love how memory management and observability is built-in.

Today, I just re-implemented Google's AlphaEvolve for Perlin noises using my own minimal agentic framework (https://toolkami.com/alphaevolve-toolkami-style/), will probably steal a trick or two from you.

Your readme doesn't mention any example of multiple agents. We don't need another wrapper to anthropic/openai/*. Without being able to select between multiple tools/agents - you only need 2 lines of code to achieve llm calls, and another 2 lines for state management.

That said the interface is more idiomatic that others agent frameworks that show up here.

It also wasn't clear where the last N messages is being persisted. Db/file/adapters/?

Good luck!

I'll just leave this here: https://github.com/The-Pocket/PocketFlow 100 lines of code agent lib that's not tied to a commercial offering