The Open Production Agent Harness for Python
Parallel tool batches, structured observations, signature loop detection, MCP tools, memory, workspace files, subagents, background tasks, and REST/SSE serving.
What It Is - Quick Start - Choose Your Path - Use Cases - Why It Matters - Install - Cookbook - Features - Docs - Ask AI
An LLM is not an agent by itself. The model provides intelligence; the harness gives that intelligence a working environment.
OmniCoreAgent is the application-facing harness layer around a model:
model
+ prompt contract
+ reasoning loop
+ local tools
+ MCP tools
+ parallel tool batches
+ structured observations
+ memory
+ context control
+ workspace files
+ tool-output offloading
+ guardrails
+ events
+ subagents
+ background tasks
+ REST/SSE serving
That is the difference between an agent harness and a generic agent library. A library gives you pieces to assemble. A harness gives you the runtime boundary that makes a model usable inside an application.
OmniCoreAgent keeps that boundary explicit:
| Layer | What It Owns |
|---|---|
| Agent harness | Model loop, prompt contract, tools, observations, memory, context, workspace, guardrails, events, subagents |
| Serving boundary | OmniServe REST/SSE APIs, request lifecycle, readiness, auth, rate limits, metrics |
| Background boundary | Durable scheduled/manual task execution with task state, run history, leases, retries, and workspace output |
| External tool boundary | MCP server tools and local Python tools exposed through one runtime surface |
Start with the core harness. Turn on heavier production pieces only when the workload needs them.
If you prefer guided docs, start with the
Quick Start.
If you use AI coding tools, use the
AI tools guide
for Ask AI, /llms.txt, hosted docs MCP, Cursor, VS Code, ChatGPT, Claude, and
Perplexity.
pip install omnicoreagentexport LLM_API_KEY=your_api_keyimport asyncio
from omnicoreagent import OmniCoreAgent
agent = OmniCoreAgent(
name="assistant",
system_instruction="You are a helpful assistant.",
model_config={"provider": "openai", "model": "gpt-4o"},
)
async def main():
result = await agent.run(
"Research the top 3 open-source agent runtimes and summarize them.",
session_id="quickstart",
)
print(result["response"])
await agent.cleanup()
asyncio.run(main())That is the smallest path: one agent, one model, one stable session, the harness loop, session memory, guardrails, workspace files, error handling, and metrics around each run.
Context management, tool output offloading, BM25 tool retrieval, subagents, skills, cloud workspace storage, and production backends are opt-in so a small agent stays small.
Ready to go deeper? The Cookbook has progressive examples from hello world to production deployments.
| Goal | Start Here |
|---|---|
| Build your first agent | Quick Start |
| Add Python tools | Local tools cookbook |
| Connect MCP server tools | MCP tools cookbook |
| Manage memory and context | Getting started cookbook |
| Save files, artifacts, and large tool results | Tool offload cookbook |
| Build a production-shaped app harness | Real applications cookbook |
| Build multi-step workflows | Workflows cookbook |
| Serve an agent over HTTP/SSE | OmniServe cookbook |
| Use the docs inside AI tools | AI tools guide |
| Debug setup or configuration | Configuration guide |
| Understand the runtime internals | Implementation Map |
OmniCoreAgent is for application builders who need the agent runtime to hold together after the prototype works.
| Build | Harness Pieces You Use |
|---|---|
| MCP-connected product agents | MCP tools, local tools, structured observations, guardrails, session memory |
| Research and analysis agents | Parallel tool batches, workspace files, tool offloading, context management, artifact readback |
| Long-running worker agents | Background tasks, durable task stores, run history, workspace output, retries, cancellation |
| Multi-agent task systems | Dynamic subagents, shared workspace output, workflow orchestration, telemetry events |
| Agent APIs | OmniServe REST/SSE, readiness, auth, request timeout, rate limits, metrics |
| Production app integrations | Optional Redis, MongoDB, SQL, S3, and R2 backends without making the core install heavy |
The core idea is simple: one harness entry point, many application membranes. You bring the domain instructions, tools, and business logic. OmniCoreAgent provides the execution boundary around them.
Most demos stop at "LLM plus tool loop." Production agents fail in the layer around that loop: slow sequential tool calls, noisy observations, repeated actions, context exhaustion, unsafe tool output, missing workspace state, uninspectable background work, and weak serving boundaries.
OmniCoreAgent exists for that layer.
The usual tool loop looks like this:
LLM -> call tool A -> wait -> result -> LLM -> call tool B -> wait -> result
OmniCoreAgent lets the model request independent tools together:
LLM -> [tool A + tool B + tool C in parallel] -> one structured observation -> LLM
The model gets one complete view of the batch before it reasons again. A failed tool is represented beside the successful tools instead of silently collapsing the whole step.
Native function calling alone is not the runtime. OmniCoreAgent uses its own tool-call contract, parser, resolver, parallel runner, and result formatter so the harness controls the full execution path.
Raw tool output is often too noisy for the next reasoning step. Large payloads, errors, irrelevant fields, and prompt-injection content can all distort the loop.
OmniCoreAgent routes tool results through an observation pipeline:
tool output -> parse -> format -> guardrail check -> offload when configured -> observation -> model
The model receives the signal it needs to continue the task, not an unbounded dump of every byte returned by a tool. When tool offloading is enabled, large outputs are written into the active workspace and the model receives a readable preview plus a path it can use later.
max_steps is still useful, but it is a blunt instrument. It stops an agent that
is making progress just as quickly as one that is stuck.
OmniCoreAgent tracks SHA256-backed tool-call signatures across the loop. Each signature is based on the tool name, input, and output for the call. The runtime detects:
- Consecutive loops: the same tool call returns the same result repeatedly.
- Pattern loops: the same tool repeats a small interaction pattern.
When the harness stops a loop, the agent gets a reason. That makes debugging the agent behavior much easier than "max iterations reached."
OmniCoreAgent ships as a working harness, not a bag of disconnected pieces:
model + prompt + loop + tools + memory + context + workspace + guardrails + telemetry
Keep it small for simple agents, then turn on the heavier harness pieces when the workload needs them: MCP tools, BM25 tool retrieval, dynamic subagents, skills, cloud workspace storage, Redis/Postgres/MongoDB memory, telemetry events, and OmniServe.
When context management is enabled, OmniCoreAgent checks the active message history before every LLM request. If the configured threshold is crossed, the harness automatically applies the selected strategy before calling the model:
messages -> threshold check -> truncate or summarize+truncate -> LLM
The system prompt is preserved, recent messages are preserved, and older middle history is either summarized or removed depending on configuration. If you set the budget below your model's real context window, the harness acts before the provider rejects the request.
import asyncio
from omnicoreagent import MemoryRouter, OmniCoreAgent, ToolRegistry
tools = ToolRegistry()
@tools.register_tool("search_web")
def search_web(query: str) -> dict:
"""Search the web for information."""
return {"results": [f"Result for: {query}"]}
@tools.register_tool("fetch_document")
def fetch_document(path: str) -> dict:
"""Fetch a domain document from an application-owned source."""
return {"path": path, "content": f"Contents of {path}"}
agent = OmniCoreAgent(
name="research-agent",
system_instruction=(
"You are a research assistant. Use tools in parallel when the calls are "
"independent and you can reason over the results together."
),
model_config={"provider": "openai", "model": "gpt-4o"},
local_tools=tools,
memory_router=MemoryRouter("in_memory"),
agent_config={
"max_steps": 20,
"context_management": {"enabled": True},
"tool_offload": {"enabled": True},
"enable_subagents": True,
"enable_advanced_tool_use": True,
},
)
async def main():
result = await agent.run(
"Search for recent AI agent papers and fetch notes.md. Do both at once "
"if neither depends on the other."
)
print(result["response"])
await agent.cleanup()
asyncio.run(main())The runtime accepts search_web and fetch_document in the same batch, returns both
results together, and continues from one structured observation.
pip install omnicoreagent # Core runtime
pip install "omnicoreagent[redis]" # Redis memory backend
pip install "omnicoreagent[postgres]" # PostgreSQL / SQL memory
pip install "omnicoreagent[mongodb]" # MongoDB memory
pip install "omnicoreagent[s3]" # S3 / R2 workspace storage
pip install "omnicoreagent[serve]" # OmniServe REST/SSE API
pip install "omnicoreagent[tokenizer]" # Token-aware context budgeting
pip install "omnicoreagent[otel]" # OTLP trace export
pip install "omnicoreagent[langsmith]" # LangSmith trace export
pip install "omnicoreagent[opik]" # Comet Opik trace export
pip install "omnicoreagent[all]" # EverythingProduction backends are installable extras. Install only what the agent actually uses.
| Feature | What It Does |
|---|---|
| Parallel Batch Tool Execution | Executes independent tool calls concurrently and returns one combined observation to the model. |
| Structured Observation Pipeline | Parses, formats, guardrail-checks, and offloads tool results when configured before the model sees them. |
| Signature-Based Loop Detection | Detects repeated SHA256-backed tool-call signatures and repeated tool interaction patterns beyond step-count exhaustion. |
| Local Tool Registry | Registers Python functions as tools with inferred schemas and async/sync execution support. |
| Multi-Tier Memory | Uses in-memory, Redis, MongoDB, or SQL-backed session history through the memory router. |
| Context Engineering | Checks context before each model call and automatically truncates or summarizes when the configured budget threshold is crossed. |
| Workspace Files | Gives agents a local, S3, or R2-backed file workspace for notes, scratchpads, artifacts, and tool offloads. |
| Tool Output Offloading | Writes large tool results to workspace files and gives the model a preview plus a file reference. |
| Guardrails | Adds prompt-injection screening inside the observation path with configurable behavior. |
| Feature | What It Does |
|---|---|
| Dynamic Subagents | Lets the main agent spawn focused workers with isolated context and shared workspace output. |
| Durable Background Tasks | Runs manual or scheduled agent work with task state, run history, retries, cancellation, and workspace output. |
| Workflow Orchestration | Provides sequential, parallel, and router agents for multi-step application workflows. |
| Telemetry and Traces | Emits typed telemetry events, retrieves traces by exact trace_id, latest session, or run_id correlation, and exports traces to OTLP, LangSmith, Opik, or JSONL. |
| OmniServe | Turns an agent into a REST/SSE service with lifecycle management, auth, rate limits, telemetry APIs, background APIs, and metrics. |
| Feature | What It Does |
|---|---|
| MCP Native Tools | Connects MCP servers over stdio, SSE, and Streamable HTTP, including OAuth-capable remote servers. |
| Agent Skills | Loads packaged capabilities implemented with Python, Bash, or Node.js. |
| BM25 Tool Retrieval | Selects relevant tools from large tool sets so the prompt stays focused. |
| Runtime Backend Switching | Switches memory backends at runtime when configured. |
| Universal Models | Supports OpenAI, Anthropic, Gemini, Groq, Ollama, DeepSeek, Mistral, OpenRouter, Azure, and Cencori through the runtime model layer. |
OmniCoreAgent's capabilities are backed by concrete runtime modules:
| Capability | Where It Lives |
|---|---|
| Parallel tool batches | core/tools/tool_batch_runner.py |
| XML tool-call contract | core/agents/xml_parser.py |
| Structured observations | core/tools/tool_observation.py |
| Tool output offloading | core/workspace/artifacts.py |
| Automatic context control | core/agents/llm_step.py, core/context_manager.py |
| Workspace files | core/workspace/tools.py, core/workspace/storage.py |
| Dynamic subagents | core/subagents.py |
| Loop detection | core/agents/loop_detection.py |
| MCP server tools | mcp_clients_connection/client.py |
| OmniServe | serve/ |
See the Agent Harness docs for the full implementation map.
All examples live in the Cookbook and are organized by use case.
| Category | What You'll Build |
|---|---|
| Getting Started | First agent, tools, memory, telemetry events, and traces |
| Real Applications | Due diligence, support operations, and workspace code review harnesses |
| Workflows | Sequential, Parallel, Router agents |
| Background Agents | Scheduled autonomous tasks |
| Production | Guardrails, serving, and production patterns |
For the first run, most hosted model providers only need LLM_API_KEY.
OmniCoreAgent defaults memory and events to in-memory storage, workspace files to
local disk, and optional production integrations stay off until you configure
them.
export LLM_API_KEY=your_api_keyAdd backend-specific variables only when you opt into Redis, MongoDB, SQL database storage, S3, R2, or OmniServe deployment settings.
The defaults keep the first agent small: workspace files and guardrails are on, conversation memory is in-memory, and advanced harness pieces stay off until you enable them. This example shows the production-style switches together.
agent_config = {
"max_steps": 15,
"tool_call_timeout": 30,
"request_limit": 0, # 0 = unlimited
"total_tokens_limit": 0, # 0 = unlimited
"memory_config": {
"mode": "sliding_window",
"value": 10000,
"summary": {"enabled": False},
},
"enable_workspace_files": True, # Default on
"guardrail_mode": "full", # Default
"context_management": {"enabled": True}, # Default off
"tool_offload": {"enabled": True}, # Default off
"enable_advanced_tool_use": True, # Default off
"enable_subagents": True, # Default off
"enable_agent_skills": True, # Default off
}When enable_subagents is true, workspace files are enabled automatically so
subagents write outputs, notes, todos, and artifacts into the active
workspace.
Full reference: Configuration Guide
git clone https://github.com/omnirexflora-labs/omnicoreagent.git
cd omnicoreagent
uv venv && source .venv/bin/activate
uv sync --dev
pytest tests/ -v
pytest tests/ --cov=src --cov-report=term-missing| Error | Fix |
|---|---|
Invalid API key |
Export LLM_API_KEY with the key for the provider selected in model_config. |
ModuleNotFoundError for Redis / Postgres / MongoDB / S3 |
Install the matching extra, for example pip install "omnicoreagent[redis]". |
Redis connection failed |
Start Redis or use MemoryRouter("in_memory"). |
MCP connection refused |
Ensure the MCP server is running before starting the agent. |
More help: Basic Usage Guide
git clone https://github.com/omnirexflora-labs/omnicoreagent.git
cd omnicoreagent
uv venv && source .venv/bin/activate
uv sync --dev
pre-commit installSee CONTRIBUTING.md for guidelines. PRs are welcome.
MIT - see LICENSE.
Built by Abiola Adeshina.
- GitHub: @Abiorh001
- X (Twitter): @abiorhmangana
- Email: abiolaadedayo1993@gmail.com
| Project | Description |
|---|---|
| OmniMemory | Self-evolving memory for autonomous agents |
| OmniCoreAgent | Production agent harness (this project) |
| OmniDaemon | Event-driven runtime for running agents as supervised, autonomous infrastructure services |
LiteLLM - FastAPI - Redis - Pydantic
Star on GitHub - Report Bug - Request Feature - Documentation