Recent advances in large language models (LLMs) like GPT-3 and ChatGPT have enabled the development of intelligent conversational agents that can understand natural language instructions and complete tasks automatically. However, building practical agents that can plan, act, and interact in real-world environments remains challenging.
To make the development and deployment of autonomous language agents more accessible, researchers from AIWaves have open-sourced AGENTS – an end-to-end framework for building LLM-powered agents and publish their results in a paper.
Some key capabilities provided by AGENTS include:
- Long-short term memory – Agents can maintain long-term memory using a database and short-term memory via a scratchpad. This allows them to track context and state over time.
- Tool usage & web navigation – Agents can leverage external APIs and search the web for information. This expands their capabilities beyond just conversational interactions.
- Multi-agent communication – AGENTS supports building systems with multiple collaborating and competing agents. A controller agent acts as a moderator and schedules actions.
- Human-agent interaction – Humans can play the role of an agent and seamlessly interact with other agents. This enables many interesting applications.
- Controllability via symbolic plans – Agents can be given high-level plans that specify states and transition rules. This makes their behavior more predictable and optimized.
In addition, AGENTS simplifies agent development by defining key components like agents, plans, and environment via a configuration file. It also enables deploying agents as APIs.
The researchers demonstrated AGENTS capabilities via case studies of single and multi-agent systems. These included a customer service agent, sales bot, fiction writing studio with multiple agents, and debate scenarios with human-agent interaction.
Overall, AGENTS provides researchers and developers an easy way to build real-world language agents with planning, tool usage, interaction capabilities and control via symbolic plans. It could expand the development and adoption of intelligent agents across many domains. The code and documentation are available on GitHub.