神经元AI代理网络系统

Neuro SAN Data-Driven Agents

Neuro AI system of agent networks (Neuro SAN) is a library for building data-driven multi-agent networks which can be run as a library, or served up via an HTTP server.

Motivation: People come with all their hopes and dreams to lay them at the altar of a single LLM/agent expecting it to do the most complex tasks. This often fails because the scope is often too big for a single LLM to handle. People expect the equivalent of an adult PhD to be at their disposal, but what you really get is a high-school intern.

Solution: Allow these problems to be broken up into smaller pieces so that multiple LLM-enabled agents can communicate with each other to solve a single problem.

Neuro SAN agent networks can be entirely specified in a data-only HOCON file format (think: JSON with comments, among other things), enabling subject matter experts to be the authors of complex agent networks, not just programmers.

Neuro SAN agent networks can also call CodedTools (langchain or our own interface) which do things that LLMs can't on their own like: Query a web service, effectuate change via a web API, handle private data correctly, do complex math operations, copy large bits of data without error. While this aspect does require programming skills, what the savvy gain with Neuro SAN is a new way to think about your problems that involves a weave between natural language tasks that LLMs are good at and traditional computing tasks which deterministic Python code gives you.

Neuro SAN also offers:

• channels for private data (aka sly_data) that should be kept out of LLM chat streams
• LLM-provider agnosticism and extensibility of data-only-configured LLMs when new hotness arrives.
• agent-specific LLM specifications - use the right LLM for the cost/latency/context-window/data-privacy each agent needs.
• fallback LLM specifications for when your fave goes down.
• powerful debugging information for gaining insight into your mutli-agent systems.
• cloud-agnostic server-readiness at scale - run where you want
• enabling distributed agent webs that call each other to work together, wherever they are hosted.
• security-by-default - you set what private data is to be shared downstream/upstream
• Out-of-the-box support for Observability/tracing data feeds for apps like LangSmith, Arize Phoenix and HoneyHive.
• test infrastructure for your agent networks, including:
• data-driven test cases
• the ability for LLMs to test your agent networks
• an Assessor app which classifies the modes of failure for your agents, given a data-driven test case
• MCP protocol API - Every Neuro SAN server can be an MCP Server.
• per-user authorization for Agent Networks - optional implementations include: OpenFGA

Extra info about agent_cli.py

There is help to be had with --help.

By design, you cannot see all agents registered with the service from the client.

When the chat client is given a newline as input, that implies "send the message". This isn't great when you are copy/pasting multi-line input. For that there is a --first_prompt_file argument where you can specify a file to send as the first message.

You can send private data that does not go into the chat stream as a single escaped string of a JSON dictionary. For example: --sly_data "{ \"login\": \"your_login\" }"