Introducing Agent-hub: An open-source agent behavior framework for your AI

The more that I use Agents, the more opportunities I find to streamline and "App-ify" the prompting experience; AI does a much better job when decisions are not just left to chance, and the agent is provided with constraints. The problem is that the more freedom you give the agent, the more likely you will receive unexpected results (and hallucinations), while typing out the constraints for each new prompt is a pain point for me.

I find that providing lengthy, highly descriptive prompts significantly slows my workflow. Having to tell an agent the exact skills to use for each prompt is frustrating. Not remembering to tell the agent to spin up subagents for tasks is annoying. Defining the personality and communication style you want the agent to use every time you open a new context window should be automatic.

Another pain point for me is lack of durable memory across AIs and sessions. It would be very helpful for the AI to automatically capture key data about the active topic and write it to a file that other agents on your computer can access automatically on session startup.

AGENTS and CLAUDE files address some of the issues; however, there are several problems that need to be remedied by Agent-hub.

1. Making "always-on" AGENTS and CLAUDE files excessively long is expensive, both in terms of tokens and response times. This is because each prompt reads the entire file when processing your input and generating the response; even if the rules do not apply to that prompt.

2. Managing rules across agents and projects (especially on teams or at the org level) is difficult to keep in sync and typically lacks a single sourt of truth.

3. You must write all the rules yourself, with no baselines or defaults, and organize the discipline structure, which is very time-consuming and easily becomes disorganized.

4. Agents do not write to durable long-term memory. Memory stays with the context you are working in.

5. Orchestration of subagents in a workflow is a challenge, and results are often inconsistent. Long running prompts in the main agent context puts agentic work at a standstill by blocking the main thread.

6. Responses must often be longer than desired to provide context, when it would sometimes be nice to respond with a simple yes/no or a checkbox and the agent already has the necessary context preloaded.

Agent-hub addresses those problems and more.

Rule structure and easy extension. Agent-hub provides an organized file structure and easy extension capabilities to customize and add custom rules, subagents, and behaviors, all in human readable text via .md files. No programming knowledge needed to write your own rules.

Wide AI compatability Agent hub uses widely recognized keywords when prompting an agent. The invocations and interactive elements should work in any AI applications. Claude, Chat GPT and Cursor users, I got you covered.

Single Source of Truth. Agent-hub is rooted in one AGENTS.md file that you can point all of your workspaces and projects to. The subagents, rules, skills, and personalities are branched from there and are contained in a single Agents folder on your computer. Users can push all customizations to a repo and updates can be easily shared accross the organization.

Give your agent a Personality. Define the personality, behaviors, and communication rules for agents. Sick of the em dash? Tell your agent to not use it. Want less verbose responses? Simply configure it in the communications artifact.

Optimized token burn. Agent-hub directs your AI to only access and utilize subagents, rules, and skills as needed for the prompt. This prevents the agents from having to reference hundreds or thousands of lines of rules, standards, and instructions for every prompt. This optimizes your token burn and improves response times while still ensuring your requirements are applied automatically and as necessary for structured prompting.

Smart subagent invocations. Agent-hub defines the triggers for when the agent should invoke subagents. You do not have to remember to call them in your prompt. When your agent receives a prompt that it it determines it should invoke a subagent for, it happens automatically for you. You don't have to remember to invoke the skills manually, although you can if needed.

Prompt engineering, simplified. One shared policy covers every app. You do not write an AGENTS document per project. Just point your projects at the hub. Triggers, handoffs, and reply ids are already written. Context stays small: one memory read, task-only bullets downstream, and reports are capped at about 80 lines.

Durable memory. Agent-hub sets up durable memory on installation for your AI. You determine where you want the memory file to be saved, and the agent handles reading and writing to it for you. Durable memory saves important information across AIs and context windows for you in a human readable format, so that you don't have to tell a new agent what you have been working on.

Easy installation. Agent-hub is very easy to install. You can either directly download the library and drop it in your workspace on your computer, or pull it down with GIT. Once you have it downloaded, just open your agent in the root directory where you placed the Agent-hub files and tell your agent to set up the hub.

User Interface Options. App-agent provides simple options, in the form of keywords and interactives to advance through option flows. You do not have to type out long responses to prompt the agent to advance though workflows.

Estimated token use, with and without Agent-hub

Using Agent-hub does come with an additonal token cost per prompt because the agent is using additonal context, rules, skills and subagents to improve your prompting experience.

These are the same additional costs you would pay if you manually invoked skills and agents, and spent the time providing important context with every prompt. Agent hub just handles providing the front matter, durable data saving, and intelligently invoking subagents and skills as necessary for you.

This table provides token estimates between a raw prompt where you would leave all the decision making up to the agent, vs using Agent-hub to invoke well engineered prompts that return optimized and increasingly reliable responses.

* Agent hub context loads on first prompt of session

Next steps for Agent-hub

Delivery of Agent-hub v1.0. More configurations, more token optimizations, improved specialists, output contracts.

Prompt Browser application. That shell sends prompts to Cursor or Claude, streams the reply into one live HTML document, and paints a fixed catalog of elements and configurable colors to the browser.

Agent-hub is ready to download & customize now!

Get Agent-hub from GitHub: GitHub project

Or direct download: Download ZIP

Learn more about Agent-Hub: README

This is an open-source project. Feel free to contribute or fork.