What is ABI-Core?¶
ABI-Core lets you build AI agents that talk to each other, discover each other, and work together — all running inside Docker containers.
The idea¶
You write Python functions. ABI turns them into agents that:
Run as independent services (each in its own Docker container)
Find each other automatically (no hardcoded addresses)
Talk to each other (a standard messaging protocol)
Follow security rules (checked before every action)
You focus on the logic. ABI handles the rest.
What it looks like¶
# Create a project with 2 agents
abi-core create project my-system --with-semantic-layer
abi-core add agent researcher --description "Finds information"
abi-core add agent writer --description "Writes reports"
abi-core run
That’s a running multi-agent system. The researcher can find the writer and send it work — without you wiring them together manually.
The 4 pieces¶
Agents¶
Python programs that use AI models to do work. Each agent has:
Steps — functions that run in a fixed order you define
Tasks — orchestrators that compose steps into workflows
Tools — functions the AI can decide to call when needed
@agent.step(name="analyze")
async def analyze(text):
result = await invoke(config.LLM_CONFIG, f"Analyze: {text}")
return {"analysis": result}
Semantic Layer¶
A search engine for agents. Instead of hardcoding addresses, you search by what you need:
agent = await tool_find_agent("someone who can write reports")
# Returns the writer agent with its address and capabilities
Under the hood it uses a vector database (Weaviate) to match by meaning, not exact words.
A2A Protocol¶
How agents talk to each other. A standard message format so any agent can call any other agent the same way:
async for chunk in agent_connection(my_card, target_card, payload):
# streaming response from the other agent
Guardian¶
The security gate. Every request can be checked against rules before it runs. Blocks unauthorized actions, logs everything.
How it’s different from just using LangChain¶
Raw LangChain |
ABI-Core |
|
|---|---|---|
Execution |
AI decides order |
Your code decides order |
Multi-agent |
You wire it yourself |
Agents find each other automatically |
Deployment |
You figure it out |
|
Security |
Nothing built-in |
Built-in rules engine |
Discovery |
Hardcoded addresses |
Search by capability |