Fury Agents Manifesto
Starting date: 21st October, 2023
ChainFury’s first commit was on 7th April, 2023. It has been about 6 months since then and it has undergone lot of production usage. With multiple API changes and engines, we are now at a stable place. This is also a good time to check up on the things that have released in the wild till now.
tl;dr
Predictable automated chains as agents, that use tree searching algorithms to find solution to a problem with given set of actions. Has ability to create new actions and learn from feedback.
Agents
There have been several “agent like” systems that have been released. Some can create code, others can perform advanced
searching. Ultimately all of them can be modelled as a Chain and use different algorithms. chainfury can support
all algorithms and has a type-based robust chaining engine. This means building agents is the next logical step. There
is a lot of theory and academic research done on the topic of agents. All of them have different tradeoffs. But first
let’s start with the requirements of an agent.
Agent should be able to execute task without human intervention
Agent should stop when it can’t proceed
Agent should be interruptible to take in feedback
Agent should take inputs from it’s environment
Agent should be able to remember things over time
Agent should be predictable in its behaviour, debuggable
Von-Neumann machine
We are followers of the agent as a Von-Neumann machine, which means each chain has a complete I/O mechanism where
each input and output can be accessed independently. chainfury can use different memory systems like VectorDB, etc.
meaning that it can persist data over time. For the CPU analogy we have chainfury.base.Chain which models the
execution as a DAG of chainfury.base.Node objects. Each node contains the unit step of the chain. We can
parallellise and speed up executions by using chainfury.utils.threaded_map.
chainfury is already being used in production and thus with the infrastructure layer sorted we can then think about
what to build on top of it.
Automated
One of the most important things is that these agents be automated and run without human in the loop.