100% make the UI outlook. Put each Agent at an email address and allow the agents to email each-other and you. This way your agents are employees in the org and you can manage their access and monitor their activities. Let them operate on and import data and save files to cloud storage. If you standardize this interface you can offer the Agent and people can onboard the agent like they onboard anyone in teams. The systems already really good at making emails and completed emails and tasks completed via email are a great data source. Build around the most data rich environment, corporate email and code generation.
Right now you and other devs are creating these chains and filling in 'memory' and context manually for each step, but we need an agent that automatically fills the right information for a specific role from each 'reflection' need - when needed. That means that all input is treated like a project, and the project managers goal, is to fill the memory of their co-helpers with the right instructions/context, at the right time. For instance. In such a project, focus would change through the dev stages, and reflection would be different in the start of the project vs later. Each co-helper are dynamically updated with updated purpose/context as we move along a project trajectory. So, instead of manual/static goals/context given from the start, we need a dynamic timeline/stage based attention and actions. Each agent will choose and 'move along' their own project attention/memory plan, and dynamically gets necessary slices of project context, webresults, agent2agent context (special needs related to a specific helper), own goal context etc, as it fits the situation. All will be addressed at the right moment in the development of an answer. Project manager runs through a few steps while adjusting overall project memory. Lead agent, runs through their own steps, and feed local memory to agents below it, and so on, so the project communication will be optimal and effective. What reflective technique/what helpers are needed for a certain project will be decided based on a quality assurance check (how many tokens pr solution), and remembered by each, in the relevant context slice. This is imho the way to automate cognitive agent architectures. This will also assure (with proper prompts), that a project can automatically adjust to the type of task given, for optimal efficiency. We already have a 'context' field that we need to control much more finely. Each agent have current instructions/context from above, and can feed relevant instructions on to sub helpers according to the current step. Slices: Super-context, own context, slice for each team member, and necessary knowledge slices to solve that step. We need a system memory/context slice manager, that can guide the attention of all the below agent hierarchy to engage in the right thoughts/reflections at the right time. There are oc also flat non-hierarchical cooperating problem solving structures, but it's easier to start with a small hierarchy.
Autogpt can use an ask the boss feature for sure, to stop it from doing unnecessary research and tasks and get on with the code making or whatever the original tasks are.
Great stuff. I'm going through all the webinars now. I'm making some updates on my framework based on this discussion
100% make the UI outlook.
Put each Agent at an email address and allow the agents to email each-other and you.
This way your agents are employees in the org and you can manage their access and monitor their activities.
Let them operate on and import data and save files to cloud storage.
If you standardize this interface you can offer the Agent and people can onboard the agent like they onboard anyone in teams.
The systems already really good at making emails and completed emails and tasks completed via email are a great data source.
Build around the most data rich environment, corporate email and code generation.
Thanks for sharing 🙏🎈 awesome work
Is it possible to add a comment where all presenters are introduced, plus add links to the correspondent repos: LangChain, ReAct, BabyAGI. Thanks!
‘nothing lasts forever…’
- famous last words
Right now you and other devs are creating these chains and filling in 'memory' and context manually for each step, but we need an agent that automatically fills the right information for a specific role from each 'reflection' need - when needed. That means that all input is treated like a project, and the project managers goal, is to fill the memory of their co-helpers with the right instructions/context, at the right time. For instance. In such a project, focus would change through the dev stages, and reflection would be different in the start of the project vs later. Each co-helper are dynamically updated with updated purpose/context as we move along a project trajectory.
So, instead of manual/static goals/context given from the start, we need a dynamic timeline/stage based attention and actions.
Each agent will choose and 'move along' their own project attention/memory plan, and dynamically gets necessary slices of project context, webresults, agent2agent context (special needs related to a specific helper), own goal context etc, as it fits the situation. All will be addressed at the right moment in the development of an answer. Project manager runs through a few steps while adjusting overall project memory. Lead agent, runs through their own steps, and feed local memory to agents below it, and so on, so the project communication will be optimal and effective. What reflective technique/what helpers are needed for a certain project will be decided based on a quality assurance check (how many tokens pr solution), and remembered by each, in the relevant context slice. This is imho the way to automate cognitive agent architectures.
This will also assure (with proper prompts), that a project can automatically adjust to the type of task given, for optimal efficiency.
We already have a 'context' field that we need to control much more finely. Each agent have current instructions/context from above, and can feed relevant instructions on to sub helpers according to the current step. Slices: Super-context, own context, slice for each team member, and necessary knowledge slices to solve that step. We need a system memory/context slice manager, that can guide the attention of all the below agent hierarchy to engage in the right thoughts/reflections at the right time.
There are oc also flat non-hierarchical cooperating problem solving structures, but it's easier to start with a small hierarchy.
Autogpt can use an ask the boss feature for sure, to stop it from doing unnecessary research and tasks and get on with the code making or whatever the original tasks are.
Astounding guys, thanks for sharing!
what's the Google paper they mentioned like Stardew Valley?
Langchain soon will fund other researchers to open source LLMs!