§ Cockpit · The operating layer above the executors
Atlas.
A meta-agent with memory. It watches your stack, briefs you each morning, learns from your feedback, and routes work to the surface where it belongs — your Slack thread, your Cowork queue, or a fix it executes on its own.
01 What it is
Atlas is the cockpit that sits above the rest of the platform stack. It is not another executor. It does not write your schema or post your content. It watches the executors that do, holds the memory of what's working and what isn't, and decides where each new piece of work should land.
In practice, you talk to Atlas in Slack. It greets you with a morning brief that aggregates the previous day across every platform. You ask it questions and it answers from live data, not yesterday's snapshot. When something breaks, it tells you in the same thread. When you give feedback, it stores that feedback, ties it to the relevant agent, and applies it next cycle. When work needs a human call, it queues a delegation in Cowork — your decision shows up as a job in your inbox, with full context.
Everything else in the agent-tools market today ships agents. Atlas ships an operating layer. That is the difference between renting a tool and operating a system.
02 Five surfaces
Atlas runs as five distinct surfaces, each tuned to a different mode of attention. You don't need to context-switch between dashboards — every surface lives in Slack, with full audit in the internal observatory.
Brief
A single message at 07:00 local time that aggregates yesterday across every platform in your stack: what published, what was scheduled, what failed, what's queued for review.
Chat
Ask anything in the channel. Atlas answers from live data — it queries the executors directly, reads the database, and returns a thread-grounded answer with citations to the source records.
Knowledge
Drop a doc, an article, a Loom transcript, a decision memo into the knowledge channel. Atlas ingests, indexes, summarizes and recalls it on demand in future conversations. Vector memory, not file storage.
Feedback
React to any agent's output with a thumb. Atlas classifies the feedback, ties it to the right agent and run, and writes it into the agent's recall index — so the next run incorporates the correction without you re-explaining.
Watchers
A standing sweep that scans for failures, drift, cost spikes, and silent regressions across the stack. When a watcher fires, Atlas posts in-thread with diagnosis and the proposed fix — never a wall of red.
03 Four modes
Atlas operates on a permission ladder. You decide how much autonomy to grant — and the ladder is the same one Anthropic recommends for production agents.
Read
Atlas observes and reports. No writes, no changes. Useful for the first week.
Default — day 1
Advise
Atlas proposes fixes and improvements as Slack messages. You approve, it doesn't act on its own.
Standard — weeks 2–4
Propose
Atlas drafts the change as a reversible artifact (a PR, a draft post, a queued delegation) and waits for your green-light.
Engaged — month 2+
Auto-execute
For allowlisted, rate-limited classes of work, Atlas executes directly. Every action carries a one-click revert button.
Compound — quarter 2+
04 Why it matters
The reason Atlas is the moat — and not just another agent — is structural. Pull any other AI agency's stack apart and you will find isolated agents bolted to a no-code orchestrator. Atlas is built the other way around: a memory-equipped cockpit that the executors plug into.
- One agent per task, isolated
- Drift detected manually, weeks late
- No conversational memory across sessions
- Feedback collected on Sheets, applied by hand
- No-code orchestrator as ceiling
- "Improvement" = quarterly re-engagement
- Cockpit reads from every executor live
- Watchers catch drift in 30-minute cycles
- Vector memory persists across every session
- Feedback routed and re-applied automatically
- Production cloud, durable workflows, full audit
- Improvement compounds every cycle
05 How it operates
Atlas runs on a production cloud, on durable workflows that survive restarts, deploys and partial failures. There is no laptop in the loop, no Zapier in the middle, no Sheets as source of truth.
Memory. A vector index for everything you drop into the Knowledge channel, plus a structured store for every conversation, gate decision, feedback event and audit row. Atlas recalls both kinds in the same answer.
Observability. Every function Atlas runs is traced with cost, latency and error rate. The internal observatory shows you per-function spend in real time. Cost drift gets a watcher alert before the bill does.
Delegation. When work needs human judgment, Atlas writes a delegation to your Cowork queue with full context attached. You open Cowork, the job is waiting with everything Atlas saw. When you complete it, Atlas closes the loop.
Governance. Every autonomous action under Mode D is allowlisted, rate-limited, and carries a revert button. The audit log is the source of truth; the cockpit is just a window onto it.
06 Live receipts
Atlas currently operates a live reference deployment — an editorial publishing platform under Umbra Group. Everything below is verifiable in the audit log.
Reference deployment — sustained for months
- 24+ durable workflows running on a single cockpit, with per-function cost and latency traced.
- 5 surfaces live in Slack (brief, chat, knowledge, feedback, watchers). Vector recall persists across every session.
- Feedback routing writes corrections into the recall index so the next run picks them up — no manual re-prompting.
- Cowork delegations route judgment calls to a queue that lives in the operator's desktop, with full context attached.
- Watchers sweep every 30 minutes and surface failures in-thread with diagnosis and proposed fix.
Source — Internal observatory + Slack audit · 2026-05
07 What it operates
Atlas is built to operate the six Studio executors below. You don't have to deploy all six — but each one you deploy plugs into the cockpit and is governed through it.
SchemaForge
Keeps every page legible to search and AI
On-page SEOOnPage Doctor
Diagnoses and repairs on-page health continuously
Entity hubsHubForge
Builds and maintains entity hub pages
DistributionPressRelay
Distributes to every channel with recovery built in
DiscoveryPulseRadar
Scans the market and drafts the response
Weekly recapLoopback
Resurfaces what worked each week, automatically
08 Deploy in your org
Atlas is the cockpit deployed at the close of every Lighthouse Sprint. The Sprint instruments the workflows; Atlas is what keeps them coherent afterwards.
Week 1. Atlas connects to your Slack workspace and reads the existing executors. Mode A (Read). Morning brief begins on day three.
Weeks 2–4. Mode B (Advise). Watchers go live. You start reacting to feedback prompts in-thread. The cockpit begins building memory of your domain.
Month 2. Mode C (Propose). Atlas drafts reversible artifacts — PRs, draft posts, queued delegations — and waits for green-light. Cowork queue activates.
Quarter 2. Mode D (Auto-execute) on allowlisted classes only. Revert button on every action. The system starts compounding without you.
09 Pricing line
Atlas is included as the cockpit in every Lighthouse Watch tier. Watch operates the platforms deployed during your Sprint; Atlas is the layer that does that operating.
If you already run agents on your own infrastructure and only want Atlas as a cockpit on top, that's a standalone engagement starting from the Watch base. Pricing is documented in the Sprint framework.
10 Engage
Want to see Atlas operating a live stack?
We run a 30-minute walkthrough on the live reference deployment: the morning brief, the chat surface, watcher alerts, feedback routing, and a delegation handoff to Cowork — end to end on production data.