Public preview version

[kunalkushwaha]

Initial public release of AgenticGoKit - A Go framework for building multi-agent AI systems

Features

• Multi-Agent Orchestration: Collaborative, sequential, and loop execution patterns
• Memory & RAG: Built-in vector databases (pgvector) with document search
• Tool Integration: MCP protocol support for dynamic tool discovery
• CLI Scaffolding: Generate complete projects with agentcli create
• LLM Support: Azure OpenAI, OpenAI, and Ollama providers
• Configuration-Driven: TOML-based setup with environment overrides

Installation

go install github.com/kunalkushwaha/AgenticGoKit/cmd/agentcli@latest

Quick Start

agentcli create my-agents --orchestration-mode collaborative --agents 3 
cd my-agents && go mod tidy && go run . -m "Your task here"

Breaking Changes

• Repository renamed from agentflow to AgenticGoKit
• Module path: github.com/kunalkushwaha/AgenticGoKit

Notes

This is a preview release. APIs may change in future versions.

Full Changelog: Initial release

---

This discussion was created from the release Public preview version.

[onestardao]

Looks like a solid launch — Go-based agent frameworks are rare and you nailed the ergonomics (the CLI scaffolding is chef's kiss).
Curious about the memory & RAG setup though — I’ve seen a lot of agent stacks fall apart subtly there.

Specifically: once you mix document vector memory + agentic flows, you start getting weird stuff like semantic drift between hops, or agents hallucinating on stale context because of how chunks were cut.

We’ve documented 16+ of these failure modes (some of them invisible till you hit production).
Not saying this lib has those — just flagging early, 'cause fixing this after you ship is hellish.

Happy to share some of the deep-dive docs if helpful. Solid work again — Go needed this.

[kunalkushwaha]

Thanks a lot , really appreciate the kind words and the sharp observations.

You’re absolutely right about the subtle failure modes when mixing vector-based memory with agentic flows. We'd love to learn more from what you’ve seen in practice ; especially the 16+ documented failure cases you mentioned. That kind of hard-earned production insight is gold, and we're very keen to integrate those learnings into the evolving design.

That said, AgenticGoKit is intentionally scoped to the orchestration layer i.e. things like agent collaboration, workflows, and control flows. We don’t directly address vector DB issues like chunking strategies or embedding hygiene, but we do aim to make it easy to layer on things like:

• Chain-of-retrieval tracking
• Retrieval-augmented self-reflection
• Memory instrumentation or auditing

The current modular design makes these additions straightforward, and we’re already exploring them.

It would be great to fold some of those deeper retrieval practices into a best-practices doc or cookbook, especially for users who bring their own RAG stack underneath. Would love to connect if you're open to sharing some of those docs or even co-developing guidelines.

Thanks again for the generous feedback and means a lot!

[onestardao]

hey this is awesome, really love how you broke down the orchestration scope and called out chain-of-retrieval tracking & memory auditing 🚀
i’m totally game to share some of those deep-dive docs and riff on how the 16+ failure modes map back to real-world hiccups. your modular design sounds rock solid, and weaving in those best-practice guidelines would be killer for folks building on top.

let me know when you’re ready to geek out & collab. cheers! 🥂

ps. here’s the master problem map in case it helps steer the convo:
https://github.com/onestardao/WFGY/blob/main/ProblemMap/README.md

[kunalkushwaha]

Hi @onestardao , Thanks for sharing the doc. Let me go through it and lets try to setup sometime end of next week.

[onestardao]

glad it helps! if it ends up saving you some pain or giving new ideas, a star on the repo would mean a lot — no pressure at all tho. just happy to be part of the convo 🙏