Agent Swarm Conversational Flow Control

[TrickyNerdy]

Agent Swarm Conversational Flow Control

I came up with a flow control chart and design for how the Executive Agents would communicate amongst themselves, after watching the Hierarchical Autonomous Agent Swarm Pt 2 video twice, David said 6:09 that communication between agents would be probably the hardest part. So here is an outline, I use a company or organization and calendar as an analogy for the loop. The wider idea is that you could instill a single proto agent at the start, that generates the handbook at the core since that's the core of what all Executive Agents share and that proto agent spins up the entire Executive layer of the swarm and perhaps the SOB itself.

Calendar System

#1 The SOB starts Daily Operations in a series of dialogs with itself and generates the Daily Operations Report and review of the last few days and operations

#2 This is summarized by the Co-SOB who decides what to send to which Executive Agent who use this Daily message to prepare for the Daily meeting. This is essentially the daily email inside the organization

#4 Each Executive Agent have a series of internal dialogues about their Daily Operations Report, and about their own operations, their sub agent's reports and their concerns as necessary and generate a brief response, optionally they can raise private concerns to Agent resources who decided what and how to present all agent's concerns to the SOB.

#5 These reports are all sent to the secretary to combine and summarize to the SOB.

#6 The SOB then Generates the Daily meeting prompt

#6.a Each agent then generates a full report and request function calls for either Q&A dialogs with other agents, information or resource requests and allocation and generation of sub agents and their task sets

#7 The SOB review these reports and requests and approves the Executive Agent to create other agents and take actions

#8 Sub Agents then perform their tasks

#9 Sub Agents report to their respective Executive Agent

Within this Calendar things can be asynchronous of course but are generally punctuated around this "daily" loop, the Calendar is broken up into 7 day weeks, 28 day months and 3 month quarters, each period has additional subroutines for oversight and maintenance of the system.

[daveshap]

Interesting. I'm not sure I entirely agree with where you're going but it's a good start. I don't like that there's a "Co-SOB" as that defeats the point and is unnecessarily complex. Overall this diagram needs to be simpler with SOB directly above Executive Agents and a few processes connecting the two directly.

[TrickyNerdy]

Co-Sob is just a really bad name, I originally called this system the Director and Co-Director, the Co-Sob is merely a specialized sub agent or mediator that does BSHR or SPR so does the secretary and Agent resources. SOB is still directly above Executive Agents, the power hierarchy is different than the operational conversation flow here. The Agent Resources is one process connecting the Executive Agents to the SOB, its also a specialized mediator. The SOB itself would be vast and complex I felt the need to narrow down the information it receives from the Executive agents and gives to the Agents with mediators. And 6.a is the solid direction connecting the SOB and Executive Agents in my system

[daveshap]

gotcha thanks that make sense

[mfilter]

Hi, I would like to point you to an existing industry standards that imho could be used for managing the communication. It's EPCIS 2.0 https://www.gs1.org/standards/epcis which is maintained by GS1. There is an open source reference implementation at https://openepcis.io/ . One interesting feature is that agents could establish subscriptions to specific events in the event repository and that the system supports role-base access (afaik). As I'm not a software developer but only a project manager (in an governmental institute) I can't tell you the details, but if you think the idea makes sense I could bring in a software developer from my team.

Just realized that my idea does not really fit to the "flow control" topic. I'll try to find a more suitable thread and link this there as well.

[sanganisahil]

What about NPC simulation game, they are using good mechanism to make this work. They also have the challenges solved. I am noob but had this thought.

[TrickyNerdy]

I knew my design above first, before I fully understood what my thoughts and feelings were, so I wanted to draw those out, now that I have them.

Essentially the process is to define a limited scope of general agents as cognition engines and use them like methods for simple SOB AIs and Executive Agents to call. The wider idea being that some proto agent could embody the structure of the HAAS and create the SOBs, the handbook and Executive Agents completely autonomously, eliminating the need to redesign the HAAS over and over. The final structure and function is still the HAAS the way David laid out, my diagram being an implementation of it. This is the theory for optimizing the communication flow keeping in mind the limitations I understand.

I don't imagine the SOB or Executive Agents would work well or to their full potential as single dimension array layers of "AI personalities" with missions in some hierarchy the way people make up advisory boards or organizations. That seems simple and anthropomorphic, too much so. A single SOB Agent that embodies Socrates for example, really can't be captured by some single profound set of system prompt, instructions, a handbook, and some data/data retrieval. Knowing what we do about latent space activation, we would want many granular Socrates' depending on the task or direction of communication because instead if we make super agents to do everything in every direction and they pass down large final entire outputs to the other Agents, bad output results and inaccuracy would just be thrown in or would generally not sub-prompt well down the chain or through the loop.

A single agent doesn't need some profound set of system prompts, instructions, a handbook, and some data/data retrieval or each contain versions of BSHR or SPR and special function calls. We could generalize and spawn agents like method calls throughout the HAAS. Members of the SOB and Executive layer could spawn agents with an on-the-fly combination of their personality, mission and daily information to do internal dialog, BSHR or SPR, and function calls. Agents' outputs would either naturally be more distilled or mediator agents could distil and pass along communication or prompt other agents in the HAAS with a definition of communication direction on a need-to-know basis. I imagine this system having more accuracy and reliability if, instead of thinking of the agents as entire people on an advisory board in an organization hierarchy, we think of them as small cognition engines that perform the cognitive task we need and generate them more dynamically. This way creating each agent is much simpler, in that their scope is very limited.

tl;dr The HAAS is still the HAAS, but to optimize the swarm and allow for reliable interconnectivity. The SOB AIs and Executive Agents should be abstracted a bit as cognition engines.

[Princebaskar]

Special Agency Capacities for Conversational Flow Control:

Proposal for Introducing a Focus Agent in the HAAS System

I’m proposing an idea that could streamline our HAAS system’s efficiency: the introduction of a “Focus Agent.”

Purpose:
The Focus Agent is designed to manage and refine agent interactions within the HAAS system. Its primary role is to prevent lengthy and unproductive conversations between agents, ensuring a more focused and efficient workflow.

Functionality:

•	Acts as a conversation moderator, stepping in when discussions become off-topic or excessively long.
•	Balances intervention by configurable thresholds, such as time limits or message counts.
•	Serves as a mediator in disputes, using predefined rules to guide its decision-making.
•	Learns from past interactions to optimize future interventions.

Configuration:

•	The Focus Agent’s authority and intervention sensitivity can be finely tuned to fit different scenarios, ensuring it assists without dominating conversations.
•	Regular audits and feedback loops will be crucial for fine-tuning its configurations and performance.

Benefits:
By implementing a Focus Agent, we anticipate enhanced coherence and efficiency in agent interactions, leading to more streamlined decision-making processes within the HAAS system.

Here’s a conceptual configuration setup for the Focus Agent:

1.	Intervention Thresholds:
•	Time-Based: Trigger intervention after a set duration of conversation (e.g., 10 minutes).
•	Message Count: Intervene after a certain number of messages without resolution (e.g., 20 messages).
2.	Authority Level:
•	Suggestion Mode: Offers guidance without enforcing actions.
•	Enforcement Mode: Makes binding decisions when necessary.
3.	Learning and Adaptation:
•	Utilize feedback from past interventions to adjust thresholds and authority levels.
•	Regularly update its decision-making algorithms based on performance reviews.
4.	Conflict Resolution:
•	Predefined rulesets for resolving common disputes.
•	Escalation protocol for unresolved conflicts, involving higher authority when needed.
5.	Performance Audits:
•	Schedule regular performance evaluations.
•	Adjust configurations based on audit outcomes and community feedback.

These configurations should be regularly reviewed and updated based on the Focus Agent’s performance and the evolving needs of the HAAS system.

Focus Agent could be categorized as a “Facilitator Agent” within the HAAS system. Here’s why:

1.	Persistence: It operates continuously, overseeing and moderating agent interactions, which aligns with the nature of a facilitator who is always present to guide discussions.
2.	Non-enforcement: As a facilitator, it suggests and guides rather than enforces, fitting your requirement for an agent that recommends but doesn’t dictate actions.
3.	Communication Enhancement: Its primary role is to streamline and improve the efficiency of agent interactions, which is a key aspect of facilitation.
4.	Conflict Resolution Support: While it may not have the authority to enforce decisions, it can play a vital role in resolving conflicts by suggesting potential solutions.

In essence, the Focus Agent as a Facilitator Agent would act as a mediator and guide, enhancing communication flow and decision-making efficiency without directly controlling the outcomes.