Four trust domains
Atlas chat, AutomationBuilder, protected context ingestion and shared R&D services now run under separate Linux accounts and separate runtime state.
ATLAS · EXACTIUS
Stronger foundation → connected workflows
Atlas is getting more reliable, more useful and more connected.
A practical update on the infrastructure work, how teams are using Atlas now, and the next execution loops linking Creative, Performance and Data.
RELIABILITY PASS COMPLETEREAL TEAM USAGENEXT LOOPS DEFINED
Use ↑ ↓ or ← → to navigate · press P to present
01 · Recent migration
The infrastructure changed underneath, without changing how people use Atlas.
No root gateways
Each service owns its own credentials and installation. One agent can no longer silently take ownership of another agent’s auth or files.
Safer updates
Fleet updates now drain services, apply changes in order, verify each account and automatically roll back when a gate fails.
Capability checks
A manifest-driven live test suite validates every declared integration after changes and every 24 hours.
Runtime refreshed
The fleet moved to the current Hermes release and GPT-5.6-sol while preserving schedules, state, skills and partner boundaries.
72 / 72 checks passed
Final migration verification completed with zero failures across the Atlas fleet.
02 · Fix the issues people actually hit
Recent friction was turned into verified fixes.
Performance queries
Creative-asset BigQuery reads now use a stable, strictly read-only service account instead of an expiring human login.
Meeting recap fallback
If Fireflies misses a meeting, Atlas now falls back to matching Gemini notes rather than silently losing the recap.
Report attachments
AutomationBuilder can attach generated HTML reports in Slack again, with delivery now proven through the actual upload path.
Channel history
Reference-message and history lookups now route through the identity that actually has read permissions.
GitHub + skills
Atlas can reliably inspect the Exactius skill surface and act through its own governed GitHub identity.
No more guessing around access
When a source is not connected, Atlas names the exact gap and routes capability expansion back to R&D.
03 · Usage now
People are bringing Atlas real work.
45coherent teammate interactions reviewed since the reliability reset on 16 July
16teammates using Atlas or AutomationBuilder across Creative, PM, Performance and operations
41Atlas interactions
4AutomationBuilder interactions
2live market sweeps: DE + UK
04 · Where usage is deepening
From retrieval to repeatable team workflows.
Brief write-back
Eva’s CRM copy workflow
Atlas compares a working copy with the original multi-tab brief, previews the exact edits, waits for approval, writes only changed text cells and preserves wireframes. Eva said it works very well and plans to use it across email types.
PM operations
Leonor, Desirée and Claudia
Daily DE and UK status sweeps, doer-specific tagging, approval chases, action-list recovery, meeting recaps, task references and live deck edits are becoming normal operational support.
Creative support
Assets, decisions and deliverables
Teammates use Atlas to locate decks and footage, explain creative-performance evidence, draft updates, rebuild slides and resolve what is blocked across ClickUp, Slack, Drive and Figma.
AutomationBuilder
Eitan’s pre-sprint recommendation flow
Eitan tested a direct-ClickUp version of the pre-sprint recommendation skill for Omaze UK. It produced the HTML report without depending on the earlier n8n read path, and report attachment delivery is now fixed.
05 · What comes next
Build the pieces so they can line up later.
Bulk ad uploads
Give Performance a controlled path to turn ready assets into upload-ready ads, with validation and inline Slack QA before anything launches.
Creative asset chain
Combine Laurent’s resizing, reformatting, QA, naming and export skills behind one Atlas invocation and one deliberate checkpoint.
Data Agent integration
Connect Atlas’s operational context to the governed analytical layer without pretending raw BigQuery access equals semantic understanding.
Richer context ingestion
Deepen Performance, DA and partner context, fill source gaps and keep evidence, freshness and uncertainty explicit.
Self-updating documentation
Maintain a human-readable capability registry across Cowork skills, Atlas-only workflows and shared automations, with governance still to be agreed.
Omaze UK context vault
Add UK as a separate governed context boundary while preserving deliberate links to shared Exactius and Omaze knowledge.
06 · Performance execution
Bulk ad uploads, built for control before speed.
The Performance team should be able to hand Atlas a verified asset set and receive upload-ready ads without repetitive platform setup.
Not blind publishing. Atlas prepares, validates and surfaces the ads for human QA. Launch remains an explicit approved action.
Use ready assets
Take the correct exports, copy, destination URLs, campaign structure and naming context.
Generate the ad set
Create platform-ready payloads consistently across variants and placements.
Check before upload
Catch missing fields, wrong dimensions, duplicate names, broken URLs and mismatched assets.
QA inline in Slack
Show exactly what will be created so Performance can approve, correct or reject before launch.
07 · Laurent’s Creative workflow
One request. A chained workflow. One useful checkpoint.
01Resize / metaversizeCreate the required asset sizes from the approved Figma source.
02Google reformatRun conditionally when the campaign includes Google Ads placements.
03Banner QACheck dimensions, content integrity and expected variants.
04RenameApply the governed campaign and channel naming rules consistently.
05Export to DriveStream directly from Figma to the approved folder, avoiding local download friction.
HUMAN CHECKPOINTReview after naming and before export. No forced approval interruption between every internal processing step.
BUILD PATH
Test the chain on a copied Figma board, review the outputs with Laurent, then package it as an Atlas-invoked Creative workflow.
08 · How the pieces line up
Creative readiness can flow directly into Performance execution.
01Claudia confirms readinessThe PM identifies which approved assets and copy are ready to move.
02Atlas runs the Creative chainResize, conditionally reformat, QA and name the exports.
03Assets land in DriveValidated exports and naming arrive in the correct handoff folder.
04Performance tags AtlasA Performance owner asks Atlas to prepare the bulk upload from that asset set.
05Atlas builds the adsCampaign payloads, variants and validation results are prepared.
06QA inline in SlackPerformance reviews the exact ads before explicitly approving launch.
DESIGN PRINCIPLE
Each feature is useful independently. The architecture keeps their inputs and outputs compatible so they can later become one governed Creative → Performance loop.
09 · Intelligence + continuity
More context, better data meaning, documentation that stays current.
Data Agent integration
Operational context meets governed analytics
Atlas contributes the reason behind work, decisions and tests. The Data Agent contributes controlled metric retrieval and semantic definitions. The handoff must preserve grain, attribution, freshness and uncertainty.
Richer ingestion
Close the current context gaps
Bring in the delivered Performance, DA and PM source maps, deepen recurring meeting and workflow coverage, and retain explicit provenance rather than dumping raw source material.
Self-updating docs
A capability map people can trust
Atlas can reconcile Cowork skills, Atlas-only workflows and shared automations into a readable guide, flag duplicates and keep change proposals human-reviewed. The final home and governance still need discussion.
Current boundary
Broad performance claims remain qualified
Read-only BigQuery connectivity is live for bounded creative-asset questions. Broader business KPI answers wait for the semantic and operational layers to be ready.
10 · Omaze UK expansion
Connected Omaze context, without collapsing the vaults.
Partner boundary
Omaze DE vault
DE-specific operating context, evidence, workflows, source access and learnings remain governed inside the DE boundary.
- DE Creative + PM operations
- DE performance and account context
- DE evidence, provenance and access rules
CONTROLLED
SHARED
CONTEXT
SHARED
CONTEXT
New partner boundary
Omaze UK vault
UK receives its own source map, operating context, permissions and ingestion rules instead of being written into the DE vault.
- UK Creative + PM operations
- UK-specific workflows and partner sources
- Independent validation and context-bleed tests
SHARED EXACTIUS + OMAZE LAYERSCompany conventions, reusable squad methods and deliberately shared Omaze knowledge can connect both vaults through explicit rules, not automatic cross-partner leakage.
THE DIRECTION
Atlas becomes broader by combining reusable foundations with isolated partner context, not by turning every account into one undifferentiated memory store.