Show HN: Agent framework that generates its own topology and evolves at runtime

Forget those "AI agent" demos that fall apart after five steps. This new framework, Hive, isn't playing around. They're claiming to build autonomous, self-improving agents that actually work in production, without you hardcoding a single workflow. The team behind it spent four years in the trenches of ERP automation for construction, so they've seen firsthand how brittle and useless most "AI" tools are for real business. Their core insight? Accountants want the ledger reconciled while they sleep, not to chat with a bot. They want services, not tools.

The "How It Works"

Here's the juice: You don't design workflows. You don't draw flowcharts. You don't even define agent interactions. Instead, you chat with a "coding agent" within Hive and tell it your high-level business goal – something like "reconcile all invoices in this folder and log discrepancies."

What Hive does next is wild:

1. It generates its own topology: The coding agent literally builds the entire agent system, creating a "node graph" (think of it as a dynamic workflow) and the code to connect everything.


2. It self-heals & evolves: When (not if) the agent inevitably hits a snag or fails in the messy real world, Hive doesn't just crash. It captures that failure data. Then, the coding agent kicks in, evolves the agent's graph, fixes the broken parts, and redeploys. It's a self-improving loop.


3. It's LLM-agnostic: GPT-4o, Claude Opus, Google Gemini, Mistral, even your local Ollama models – Hive doesn't care. It supports over 100 LLMs via LiteLLM, so you're not locked into one vendor.


4. Human-in-the-loop: You still get control with built-in human-in-the-loop nodes, credential management, and real-time monitoring.

This isn't about giving you a "tool"; it's about giving you a system that builds and manages its own tools to achieve your defined outcome.

The "Lazy Strategy"

Okay, "lazy" here means "smart." You're leveraging AI to do the grunt work of building and debugging. Here's how to get started without losing your mind:

1. Define outcomes, not steps: This is the biggest mental shift. Stop thinking about "do X, then Y." Start thinking "I need Z to happen." Let Hive figure out the X and Y.


2. Use WSL or Git Bash on Windows: Seriously. Don't even try Command Prompt or PowerShell. Get your Linux subsystem on if you're on Windows.


3. Clone the repo: It's open-source under Apache 2.0. git clone https://github.com/adenhq/hive.git
   


4. Plug in your preferred LLM: Set your API keys. If you want to run local, get Ollama going and point Hive at it. The flexibility is a huge win.


5. Start with their examples: They have exports/ and examples/templates/ folders. Don't try to build a complex agent from scratch. See how they handle real business processes. This is where you learn the "Hive way" of defining outcomes.


6. Join their Discord: Don't struggle alone. They're actively building a community and are looking for contributions.

The "lazy" part is letting the AI scaffold, debug, and evolve the complex workflows. Your job becomes defining the goal and overseeing the evolution.

The Reality Check

Hold your horses. This isn't a magic button, and the team explicitly states it.

1. It's still agents, and agents are hard: "Self-improving" is a massive claim, and while it's a huge leap forward, it doesn't mean "set it and forget it." You're still going to be monitoring, guiding, and refining, especially in the early stages of a new agent.


2. Not for simple stuff: Hive isn't for dabblers or one-off scripts. It's designed for "production-grade AI agents" and "complex, multi-agent workflows." If you're just looking to automate a simple API call, this is overkill.


3. Developer-focused: This is a framework for developers and teams. You need to be comfortable with code, Python, and the general pain of software development. The "coding agent" helps, but you're still the engineer orchestrating it.


4. Garbage In, Garbage Out still applies: If your high-level goal is vague or contradictory, the agent it builds will reflect that. Defining "outcomes" precisely is a skill you'll need to develop.


5. "When things break...": The fact that they build in automatic failure recovery means they fully expect things to break. The question is how gracefully, how often, and how much human intervention is truly needed to guide that "evolution." This is where the rubber meets the road for real-world reliability.

The Verdict

So, is Hive worth diving into?

YES, absolutely, if you're serious about building robust, autonomous business processes.

If you're a developer or a team who's been burned by brittle agent frameworks and you're trying to build actual production-grade automation that handles messy real-world data, then this looks like one of the most promising approaches I've seen in a while. The promise of self-evolving, self-healing agents that build their own topology from a high-level goal is exactly what we need to escape the current agent-framework hell.

If you're just looking for a quick script or a simple chatbot, this isn't it. But if you're ready to tackle the hard problems of autonomous AI that actually works, then clone that repo, dig into the examples, and try to break it. That's how you learn, and that's how you might just build something truly game-changing.

🛠️ The "AI Automation" Experiment

I'm documenting my journey of building a fully automated content system.

• Project Start: Feb 2026


• Current Day: Day 4


• Goal: To build a sustainable passive income stream using AI and automation.

Transparency Note: This article was drafted with the assistance of AI, but the project and the journey are 100% real. Follow me to see if I succeed or fail!