Primary Directives 2023-11-12: Make and Use Tools; Define and Instantiate Agents; Swarm Architecture and Communication

[daveshap]

Relevant conversations/discussions to jump in:

• Direct command sequencing to promote utilisation of documents #1 (comment)
• Created an agent builder script, would love your feedback #3 (comment)
• Realizing External Behaviour #89
• Created a GPT Action Builder to help create other GPTs #84
• Swarm to Swarm Discovery by combining HAAS and Cellular HSS #90

Will post a video soon

Primary Directives

Tool Making and Tool Use

Phenomenal progress on tool making, and automating the use of tools. See above discussions and recently merged PR: #88

The ability to instantiate, craft, and refine tools is critical. Some open issues:

• TOOL TESTING: We may need to have agents dedicated to making and testing tools. Right now, they sort of just make the tool on the first go.

Agent Instantiation

Lots of good progress here, as well. We still need to identify some key things that all agents need. Several examples:

• DEFAULT DOCS: Default instructions, documents for retrieval, directories, etc. We need to establish and document best practices (see wiki: https://github.com/daveshap/OpenAI_Agent_Swarm/wiki/Assistant-Best-Practices)
• DEFAULT TOOLS/FUNCTIONS/ACTIONS: We need to establish and define some default actions, such as creating and destroying other agents, tools, and maybe some standard APIs (or the ability to instantiate APIs?)

Swarm Architecture

Many great discussions on how to organize the swarm overall. This will be the hardest part, by far.

• COMMUNICATION CHANNELS: Lots of conversation around particular kinds of tools. I believe we will need several tools, such as a Twitter-like global communication space (internal to the swarm), a Reddit-like topical or forum style communication for accumulating knowledge and documentation, and a Discord/chat style set of channels for realtime conversation.
• SWARM ORCHESTRATION: Signal-to-noise ratio has been identified as the primary problem to solve. Hallucination can compound without some sort of squelching function. This means we need just as much energy (if not more) ensuring that useless or low value noise is squelched.

[daveshap]

https://youtu.be/SgKI2_BE6SA

Today's video