Balancing human guidance and self‐sufficiency

Two parallel ways of working on HAAS have been identified. As explained on this page both are required to achieve the objectives of the project and will be worked on in parallel with cross-pollination at crucial decision points.

Human guidance

It is hard to build a successful self-sufficient agent swarm when it is unclear how an expertly crafted swarm would look like or what it would be able to achieve. For that reason work needs to be put into manually testing different:

• Agent network topologies
• System prompts
• Communication protocols (through tool function and Python logic backed in the underlying framework)

This approach focuses on generating low agency networks. Ie.: agents don’t have a say about many of the system characteristics, but are instead asked to exclusively focus on task solving. By limiting the variables at play a great deal of valuable information can be recovered.

The objective of this path is to provide a good understanding of what can be achieved with the current frontier models when they are optimally organised into larger networks with efficient communication protocols. This information will serve as a measuring bar and enable effective developments in self-sufficiency.

Self-sufficiency

To enable the expected full potential of HAAS, agents must be given control over how to organise themselves, how to build tools, and which tools to use. Thus this second approach focuses on high-agency problems, such as autonomous tool creation and self-arranging topologies.

In time this will lead to highly-capable agents, which will help reduce the size of the swarm. This has some advantages:

• Reduced noise
• Lower token consumption