# Minutes of Meeting — Enterprise Brain V2 Architecture and Agent Model
**Date:** March 3, 2026  
**Source transcripts:** [20260303.txt](20260303.txt), [20260303-2.txt](20260303-2.txt), [Mar 3, 2026-2.txt](Mar 3, 2026-2.txt)

**Attendees:** Yeshwanth Reddy Yerraguntla, Vara Kumar Jagarapu, Manisha Gundapuneedi, Rajashekar G, Satyasri Prabhakar Mantripragada, Gopal Gottumukkala

## Act I — 00:00 to 00:04, reject the trivial version

Yeshwanth started by explaining the current product in plain terms: multiple connectors, answering ability, and dynamic visualizations. He then immediately warned that the visible demo still felt too small for the platform ambition.

> *"We want to have a system that can do things by itself."*
> — Yeshwanth Reddy Yerraguntla _(~00:01:24)_

> *"I still feel we will be showing something trivial only."*
> — Yeshwanth Reddy Yerraguntla _(~00:02:39)_

That tension framed the rest of the meeting. The product had to stop being “a chatbot with charts” and become a system that acts before the user asks.

## Act II — 00:04 to 00:12, what versus how

The next block clarified the split between the intelligence that decides and the interface that renders. Yeshwanth kept returning to the same idea: the architecture must separate the what from the how.

> *"This is focusing on the what aspect."*
> — Yeshwanth Reddy Yerraguntla _(~00:28:10)_

> *"For every user I have to tell them news in their own fashion."*
> — Yeshwanth Reddy Yerraguntla _(~01:07:08)_

That led into the first version of the system structure: dashboard, chat, and alerting on top, with the dashboard acting like a personalized newspaper and the alert channel reserved for urgent items.

## Act III — 00:12 to 00:21, refresh loop and snapshot

The discussion then moved into the refresh loop. Data agents fetch from sources, the real-time agent updates the organization snapshot, and the user-aware distribution path decides what each person should see.

> *"We are calling it A2A."*
> — Yeshwanth Reddy Yerraguntla _(~00:34:50)_

> *"This will ping the data agents."*
> — Yeshwanth Reddy Yerraguntla _(~00:22:08)_

The team was careful to keep refresh-triggered work separate from ad hoc user queries. One path is background freshness; the other is a direct request. Both need to be visible in the architecture.

## Act IV — 00:21 to 00:31, broadcast and user awareness

That distinction became more concrete when the group discussed how the same update should reach different people in different forms. The snapshot stores the current state, the distribution engine decides who gets what, and the UI layer decides how to represent it.

The goal here was not only to move data around but to reduce load on the core agent. The system should not recompute user-specific delivery inside the same box that already has to synthesize the organization state.

## Act V — 00:31 to 00:42, skills, runbooks, resources

The architecture then turned to repeatability. Yeshwanth described skills, runbooks, resources, and custom indexes as the way the agent stops reinventing common queries and starts behaving consistently.

The team treated these as separate buckets because they do different jobs. Skills package repeated functionality, runbooks capture process, resources hold stable documents, and indexes preserve critical signals like sentiment or escalation state.

## Act VI — 00:42 to 00:49, caching and final shaping

The last segment connected the architecture back to performance and operational realism. Query caching and source-specific indexes should keep the system from re-running expensive logic for every repeated question.

The meeting ended with the same theme it started with: the platform has to become proactive, but it also has to stay structured enough that each agent can be reasoned about, instrumented, and extended.

## Todos

<todo>
  Define the agent repository skeleton with `agent.py`, `tools.py`, `utils.py`, and `db.py` for each core agent family.<br/>
  <span class="owner">Yeshwanth / platform team</span>
  <span class="deadline">Next implementation pass</span>
</todo>

<todo>
  Finalize the refresh loop, organization snapshot updates, and user-aware distribution path as separate responsibilities.<br/>
  <span class="owner">Core architecture team</span>
  <span class="deadline">Before the next architecture review</span>
</todo>

<todo>
  Specify the shared request context, telemetry, and logging contract that every agent must carry.<br/>
  <span class="owner">Rajashekar / Yeshwanth</span>
  <span class="deadline">Before connector implementation</span>
</todo>

<todo>
  Capture the first reusable skills, runbooks, and custom indexes for Jira, Gmail, and Salesforce.<br/>
  <span class="owner">Data source owners</span>
  <span class="deadline">During the next build cycle</span>
</todo>