What are sovereign agents
Sovereign agents are extension of a sovereign individual, expanding his capabilities and augmenting his intelect.
Why agents
Currently primary modality of how people interact with any “AI” is by requesting things from it. But this is merely a crutch compared to the endless amount of things an semi-autonomous agent could do for you.
Why sovereign agents
- technology advancements and state of the world has brought us to a point where it is possible and also needed to build them
- pushing the boundries of possible and advancing the frontier of AI and human capabilities
- creating a free(er) market for compute and data and escape the big tech surveillance bubble
Guiding principles for sovereign agents are:
- they preserve the individuals privacy
- they support the individuals goals
- they save the individuals time
- they enhance the individuals productivity
- they enable the individuals sovereignity
How sovereign agents operate
Sovereign agent is a guard dog for its owner’s attention and time. Proactively gathering and processing infromation on behalf of its owner, filtering out and summarizing the relevant and discarding the noise. And push the relevant to the owner instead of waiting for the owner to figure out it needs it.
Sovereign Agent can be thought of as a personal assistant knowing all your quirks, executing your will and making sure you’re on top of everything needed. It should improve your life, not provide just another gadget to play with.
Functions that sovereign agents could do:
- news aggregation and preparation of daily summaries based on interests
- email and other communication filtering, automated responding acting as a narrow executive assistant (or at least draft responses)
- perform simple tasks
- web research
- copywriting
- market research
- write code (and execute it)
- autonomously perform (basic) tasks (at some point)
Building blocks for Sovereign Agents
Data vending machines
Data vending machines simple apps that provide data or services for payment or free. It utilizes nostr for discovery and coordination, meaning its very simple to both programatically discover services or spin up new dvm’s to offer a service. DVMs can work as connective tissue, bridging the gaps in data, knowledge and access and enable agents to purchase what they need. And at the same time enable people to offer services consumable by agents.
Model Context Protocol / Open Tool Calling
MCP was developed by Anthropic and expands LLM’s capabilities with tools. There is an implementation of using nostr as discovery mechanism for MCP called DVMCP. OpenToolCalling is trying to achieve similar goals (and is compatible with MCP).
Confidential Computing
Confidential computing enables us to run sensitive operations on much more capable hardware than we own. This enables us to securely scale while not relying on big tech solutions like OpenAI for our AI needs.
Open and interoperable approach to building
- building in the open to attract as many people to the efforts as possible
- thinking about developer experience to make new contributions as simple as possible
- defining API contracts (where possiblle) to make integrations easy(ier)
Why can’t I just run everything locally
- you could, and it should be supported to prioritize local utilization of resources, but you likely won’t be able to have the hardware resources or access to all information you’d like locally
- confidential computing and utilization of distributed compute with DVM’s gives you the ability to access large amount of information, algorihtms and capabilities that you don’t have locally
- distribution of workloads improves efficiency of second realm resource utilization and provides business opportunities for specialization
How do we build them
implementing reusable building blocks using holon philosophy, making each of them usable by themselves and as a part of a larger system
- most of holons can be implemented as standalone data vending machines, thus immediately useful and could actually provide revenue
- by embracing DVM’s sovereign agents instantly get all the abilities of all DVMs out there (with the caveat that they are not confidential)
- easily reusable confidential computing playbooks and prepared containers for developers to start creating software that can run in enclaves
focusing upfront on making the participation of other interested parties as smooth as possible
- extensive documentation and development tooling to enable rapid experimentation
- ensuring minimal non-breakable communication standards between holons to prevent breaking changes and friction in the process
- making scratching your own its trivially simple
embracing our weakness
- sovereign agents are not a commercial product but an implementation proposal of distributed agents for sovereign individuals and thus don’t care about controling every aspect of the experience but instead embracing chaos
- horizontal distribution and interoperability can far outweight lack of vertical integaration on the long run
- open systems tend to win over closed ones given enough time to catch up on technological advancements
currently identified obstacles
Privacy
- confidential computing is complex, with multiple competing architectures and implementation limitations (more research on that can be found at ungovernable.tech
- limited availability of hardware, specifically high end GPUs that have secure enclave support but there are alternative approaches like extending CPU TEEs on existing GPUs without hardware modifications
Complexity
- lots of new technologies that might lack maturity and will be bumpy to work with
- decentralized systems are inherintly more complex to build than centralized ones
Resources
For comments, issues or ideas go to repository or ping me on nostr.