Brain AI · Atlas Context POC

Shared Context Structure

A growth operating system runs on context, not raw documents. This is the three-layer model agreed in the David design sync: a stable Exactius company foundation at the base, the partner's own fundamentals in the middle, and a live partner operational layer on top that ingestion keeps current.

First partner scope: Omaze DE Context agent: Atlas Operating system: Brain AI Source: design sync 2026-06-25
Exactius company — stable Partner fundamentals — semi-stable Partner live ops — changes daily Ingestion / feeds
L2 L2 · LIVE OPERATIONS updates ~continuously

Partner live operational context — Omaze DE

Partner-specific, but the fast-moving layer: ongoing initiatives, operations, and learnings. Updated continuously by ingestion, and what Atlas answers from day to day.
Ongoing initiatives & active work Live project / task state Learning agenda & learnings Testing roadmap in flight Decisions log Squad status Creative flavors in flight
L1 L1 · PARTNER FUNDAMENTALS updates ~slowly

Partner / customer context — who they are & how they work

The customer's own world: their goals, product, and the way their work runs. Semi-stable, changes when the partner's strategy or structure shifts.
North star & KPIs OKRs & goals Product & brand Customer journeys & funnel Product / squad workflows Tracking plans & metric meanings Data layer & sources
L0 L0 · EXACTIUS FOUNDATION ~static

Exactius — the company & how it operates

The bedrock: Exactius as a business and agency. What it does, how it does it, and the operating model every partner engagement runs on. Rarely changes.
What Exactius does (business overview) AI-first squad model Role expectations & how squads run Playbooks & repeatable work Standards & ways of working System map / tooling

What updates the context

Slack
Project / shared channels (not DMs) — operational chatter, decisions.
ClickUp
Tasks, comments, status, lane / partner fields — operational state.
Meeting transcripts
Fireflies / Gemini notes from meetings Atlas is invited to.
Drive / docs
Narrative briefs, decks, workflow boards, role docs.
GitHub
Repos, issues, PRs, and data-wiki content owned by analysts and engineers.
Other sources
Additional surfaces as squads define them — design, analytics, and partner tools Atlas is granted.
Manual handoff
Human-in-the-loop submissions for what auto-ingest misses.

What the context feeds

Atlas answers
Status questions, asset progress, "what to test next & why".
Learning agenda
What we've learned per house / partner, linked to OKRs.
Testing roadmap
Next tests, the lane they belong to, the goal they connect to.
Notifications
Operational nudges, exports, status surfacing.
Workflow automation
End-to-end flows — brief to task to assignment to status update, triggered from context.
Task execution
Create / update / route ClickUp tasks, request exports, kick off recurring runs.
Cross-squad handoffs
Auto-route work across Perf / Creative / Data junctions with the right context attached.
Proactive alerts
Flag blockers, missing assets, launch-crunch SLAs and stale items before they bite.
Custom agents
The foundation for purpose-built agents that execute work across the Exactius organisation.
Weekly status deck
Atlas added to the deck for ongoing partner reporting.

How it flows — the ingestion cycle

Not a one-way pipe. Curated context carries the ingestion rules that govern it, and the review step feeds corrections back into those rules, so each turn of the loop gets sharper.

↻ Ingestion cycle review tunes the rules
01 · SOURCES
Raw inputs

Slack, ClickUp, transcripts, Drive, GitHub — only surfaces Atlas is granted.

02 · ATLAS INGEST
Hermes crons

Scheduled ingestion applies the current rules — exclusions, confidentiality, priority.

03 · CURATE + RULES
Curated context

Normalised into the three-layer model. Ingestion rules live here: what to pull, how to bucket, what to drop.

04 · REVIEW & ADJUST
Human-in-the-loop

Owners check quality & relevance, fix wrong calls, tune buckets and exclusions.

05 · SERVE
Outputs

Learning agenda, testing roadmap, status answers, automations — with source evidence.

↻

The loop: step 04 feeds straight back into the ingestion rules at step 03. Every correction sharpens what gets pulled, how it is bucketed, and what is excluded next turn, so curated context keeps improving instead of drifting.