abi-core-ai 1.9.2

Agent-Based Infrastructure Core - Runtime and CLI

ABI Swarm 🤖

ABI Swarm is a self-building multi-agent system. It plans, orchestrates, and creates agents on demand — governed by policy at every step.

ABI-Core is the framework underneath: the foundation for building Agent-Based Infrastructure where intelligent agents collaborate through semantic context, policy-driven governance, and modular orchestration.

Beta Release — APIs may change. Feedback welcome.

---

What Makes ABI Different

Most agent frameworks assume your agents and tools already exist. ABI doesn't.

• The Orchestrator triages requests (simple vs complex) and coordinates execution
• The Guardian validates every query for prompt injection and policy compliance before it reaches any agent
• The Planner decomposes complex requests into tasks using Tree of Thoughts reasoning
• The Builder creates ephemeral agents on demand — Docker containers that live, execute, and die
• The Semantic Layer connects everything through vector search, MCP tools, and artifact storage
• The MinIO Artifact Store provides object storage for code, data, and results between agents

If an agent doesn't exist for a task, ABI builds one. If a tool doesn't exist, ABI creates it. All governed by policy.

---

Quick Start

pip install abi-core-ai

# Create a complete swarm (project + all agents + services)
abi-core create swarm --name my-swarm

# Or step by step:
abi-core create project my-system --with-semantic-layer --with-guardian
abi-core add abi-swarm

# Run
abi-core run              # Detached (banner + status only)
abi-core run --logs       # With container logs
abi-core run --build      # Rebuild containers first

---

Architecture

User Request
     │
     ▼
┌──────────────────────────────────────────────────┐
│                  Orchestrator                     │
│                                                   │
│  ┌─────────────┐  ┌──────────────────┐           │
│  │  Triage     │  │  Guardian Gate   │  Level 0  │
│  │  (simple/   │  │  (security       │  parallel │
│  │   complex)  │  │   validation)    │           │
│  └──────┬──────┘  └────────┬─────────┘           │
│         └────────┬─────────┘                      │
│                  ▼                                 │
│         ┌────────────────┐                        │
│         │ Gate Decision  │  Level 1               │
│         │ approve/block/ │  merge                 │
│         │ respond/plan   │                        │
│         └───────┬────────┘                        │
│                 │                                  │
│    ┌────────────┴────────────┐                    │
│    ▼                         ▼                    │
│  Simple                   Complex                 │
│  → LLM responds           → Planner → Builder    │
│    directly                 → Execute → Synthesize│
└──────────────────────────────────────────────────┘
         │                         │
         ▼                         ▼
┌─────────────┐  ┌─────────────┐  ┌─────────────┐
│  Semantic   │  │  Guardian   │  │   MinIO      │
│   Layer     │  │  Security   │  │  Artifacts   │
│             │  │             │  │              │
│ Agent Cards │  │ OPA Policies│  │ Code, Data   │
│ Tool Cards  │  │ A2A Auth    │  │ Results      │
│ Service Cards│ │ Audit Log   │  │ S3-compat    │
│ MCP Tools   │  │ Risk Score  │  │              │
└─────────────┘  └─────────────┘  └──────────────┘

---

Core Concepts

AbiCore — FastAPI-style Agent Runner

from abi_core.agent import AbiCore
from my_agent import MyAgent

agent = AbiCore()

@agent.task(name="step1")
def step1(raw_input):
    return {"cleaned": raw_input.strip()}

@agent.task(name="step2", depends_on=["step1"], input_map={"data": "$step1.result"})
def step2(data):
    return {"stored": True}

@agent.tool(name="search")
def search(query):
    """Available to the LLM on demand."""
    return {"results": []}

@agent.mcp_tool(name="bigquery_search", input_map={"query": "$input.user_query"})

agent.run(MyAgent())

invoke() — Unified LLM Calls

from abi_core.agent import invoke

# LLM only, no context
result = await invoke(config.LLM_CONFIG, "Classify this query...")

# LLM with conversation memory
result = await invoke(config.LLM_CONFIG, "Summarize...", thread_id=context_id)

# Agent with tools, no context
result = await invoke(config.LLM_CONFIG, "Find agent...", tools=[tool_find_agent])

# Agent with tools and memory
result = await invoke(config.LLM_CONFIG, query, tools=[find, search], thread_id=session_id)

ToolExecutionGraph — Deterministic DAG with Parallelism

Tasks registered with @agent.task() are wired into a LangGraph DAG. Nodes at the same dependency level execute in parallel. The LLM never decides execution order — the graph does. Retry, checkpoint/resume, and $reference resolution between nodes are built in.

Three Types of Cards

Card	Purpose	Example
AgentCard	Identity for agents (permanent or ephemeral)	Planner, Builder, zombie agents
ServiceCard	Identity for non-agent services	Webapp, semantic layer
ToolCard	Metadata for MCP tools with access_scope	query_sales_db, send_email

All cards support HMAC authentication, checksum verification, and OPA policy evaluation.

Zombie Container Pattern

Ephemeral agents run a 3-phase DAG:

1. gather_context — Pull artifacts from MinIO, prepare workspace
2. analyze_and_execute — LLM autonomous with tools (decides what to call)
3. synthesize_and_report — Package results, upload artifacts

No docker build — just docker run with env vars. The zombie self-configures as a fully functional A2A agent.

---

Features

• AbiCore Runner — agent = AbiCore() with auto-config import
• Decorator API — @agent.task(), @agent.tool(), @agent.mcp_tool()
• ToolExecutionGraph — LangGraph DAG with parallel execution and Annotated reducers
• invoke() — Unified LLM calls: LLM-only, agent+tools, with/without context
• SSE Heartbeat — Automatic keepalive for CloudFront/proxy compatibility
• Session Management — Auto process_answer(), clear_session(), _yield_clarification() in AbiAgent base
• AgentResponse — Typed responses: .success(), .error(), .status(), .input_required()
• A2AResponse — Streaming event parser for TaskArtifactUpdateEvent and TaskStatusUpdateEvent
• Multi-Provider LLM — Ollama, OpenAI, Anthropic, Bedrock, Azure, Vertex AI, Grok
• MCP Protocol — Streamable HTTP transport with auto-reconnection
• MCPToolkit — Dynamic tool calling with ServiceCard or AgentCard auth
• Three Card Types — AgentCard, ServiceCard, ToolCard with HMAC, checksum, access_scope
• Guardian Gate — Parallel triage + security validation before any agent executes
• OPA Policies — Support for agent://, service://, tool:// prefixes and ephemeral agents
• Semantic Layer — Weaviate vector search for agents, tools, and services
• ArtifactStore — S3-compatible object storage (MinIO local, AWS S3, GCS production)
• ContainerRuntime — Docker lifecycle abstraction (run, destroy, health check)
• Builder — Creates ephemeral agents with register/deregister lifecycle
• Zombie Pattern — 3-phase DAG: gather → execute (LLM autonomous) → report
• Tree of Thoughts — Planner and Orchestrator prompts with multi-path reasoning
• Infrastructure Filter — Prevents builder/planner/orchestrator from being assigned as task executors
• CLI — abi-core create swarm, abi-core create project, abi-core add abi-swarm, abi-core run

---

CLI Commands

# Swarm (everything in one command)
abi-core create swarm --name <name>

# Project (step by step)
abi-core create project <name> [--with-semantic-layer] [--with-guardian]
abi-core add abi-swarm

# Run
abi-core run              # Detached (banner + status table)
abi-core run --logs       # With container logs
abi-core run --build      # Rebuild containers

# Agents
abi-core add agent <name> --description "..."
abi-core remove agent <name>

# Services
abi-core add semantic-layer
abi-core add service guardian-native

---

Project Structure

my-swarm/
├── agents/
│   ├── planner/          # Task decomposition (Tree of Thoughts)
│   ├── orchestrator/     # Triage + Guardian gate + workflow coordination
│   ├── builder/          # Ephemeral agent creation + registration
│   └── my-agent/         # Your custom agents
├── services/
│   ├── web_api/          # FastAPI application + ServiceCard
│   │   └── service_cards/
│   ├── semantic_layer/   # MCP server + Weaviate
│   │   ├── agent_cards/
│   │   ├── tool_cards/
│   │   └── service_cards/
│   └── guardian/         # OPA security + audit
├── compose.yaml          # All services + MinIO + OPA
└── .abi/runtime.yaml

---

Documentation

https://abi-core.readthedocs.io

---

Contributing

git clone https://github.com/Joselo-zn/abi-core
cd abi-core-ai
uv sync --dev
uv run pytest

---

License

Apache 2.0 — see LICENSE

---

Built by José Luis Martínez — Creator of ABI (Agent-Based Infrastructure)