DeepEngine is the agentic workspace where AI agents execute real work across browsers, business systems and infrastructure. Every action is governed by your policies, approved by your people where it counts, and recorded in a tamper-evident audit trail.
Has the supplier confirmed our two orders from last week?
I am looking in the supplier portal.
Order 2291 shipped on Tuesday. Order 2304 still waits for us to confirm the delivery date. The portal proposes 14 March, the date you agreed on in #purchasing.
Confirmed. The portal now shows 14 March for order 2304.
@Atlas what do I have to prepare for someone who starts on Monday?
I am looking through the docs.
Three things: order the laptop from the standard list, enrol it before the first day, and book the security introduction. The VPN step is missing on “Laptop setup”, so I proposed adding it.
It becomes part of the docs once a person has reviewed it.
A new ticket: Muster AG receives every invoice twice. I am reading up on it.
The duplicates began on Monday, the day the export schedule changed. That fits the note in the docs about running both schedules for a week. Sending the reply stays with you.
Two export schedules are active. The draft explains it and names Friday as the end.
@Nightwatch the pods in auth-service keep restarting. Can you find out why?
I am checking the last deployment and the logs on prod-cluster.
The restarts began with today’s deployment. The new version expects a variable that is not set. I prepared a rollback to the previous version.
Rolled back. All three pods are running again, and the restarts have stopped.
Every agent drives a browser of its own. It authenticates with credentials from the Vault, and you can watch the session live and take control at any moment.
Talk to an agent from any room, on the web or on mobile. Long answers are condensed to a spoken summary while the full output stays in the thread.
Agents run hybrid search across your docs and propose changes for human review. Every room keeps a persistent memory of its own.
Laptop setup: add the VPN step
Connect Linear, Notion, GitHub and more in one step, or bring any MCP server or OpenAPI specification. Agents discover the right tool through search.
A lightweight worker connects Kubernetes clusters, Linux servers and Macs. On a Mac set up for it, agents operate native applications.
Trigger agents from messages, reactions, schedules and incoming events. Scheduled check-ins review open work and speak up only when someone needs to know.
Schedule: Mondays, 07:00
Create a run for Digest
Monitoring: a check fails
Create a run
Work that needs a machine will no longer need one of yours. DeepEngine Runtime gives an agent a Linux or macOS environment that we operate: a place to run commands, build, and use desktop applications, under the same policies and approvals as everything else.
An agent reads, correlates and works out the next action.
Your restrictions allow the action, hold it for approval, or deny it outright.
A sensitive action waits for a human decision: on the web, on mobile, or by voice.
The click, the call or the command executes once, and the audit trail keeps the evidence.
click “Confirm delivery date”portal.supplier.exampleDeny rules and approval rules for commands, file changes, browser clicks and calls to connected services.
Approvals reach the people authorized to decide: on the web, in the mobile app and on a call.
Workers hold no model and no model credentials. They connect outbound only and accept nothing but signed, short-lived permits.
Every run, approval and action lands in an append-only log chained by hashes, and a check verifies the chain.
Give it a name, a purpose and its instructions. Then define what it can reach, how much autonomy it has and who may use it. Each agent gets a dedicated room and a memory of its own.
How agents workResearches and completes work on websites.
InstructionsUse the supplier portal for anything about orders. Report what you found before you change something, and never pay or cancel without approval.
Answers from the team’s docs and keeps them current.
InstructionsAnswer from the docs and cite the pages you used. When a page is wrong or something is missing, propose the change instead of working around it.
Triages new tickets, finds the likely cause and drafts the reply.
InstructionsFor every ticket assigned to you: read the history, isolate the cause, and leave an internal note with a draft reply. Never send a public reply.
Investigates monitoring alerts and remediates what policy allows.
InstructionsWhen an alert arrives, establish what changed before you touch anything. Restart only stateless services yourself; request approval for everything else.
Compiles last week’s decisions and open items every Monday.
InstructionsOn Mondays, report what was decided, what is still open and who is waiting for whom. Ten lines at most, and no praise.
Agents act through their browser, through connected services and through workers on your own systems. Read-only tools run immediately. Everything else requires a human decision until you set the policy.
Also in the catalogue: Asana, Intercom, Stripe, NetBox, 1Password and Alertmanager. Any other service connects through an MCP server or an OpenAPI specification.
Requests, look-ups, checks and recurring work. Agents pick them up in the channel where your team already works.
A schedule starts Digest. It reports what was decided last week and what is still open.
A question in #purchasing has waited since Friday, so it says so. Otherwise it stays quiet.
Atlas answers from two pages and proposes the missing VPN step for review.
Scout wants to confirm a delivery date in the supplier portal. She approves it on her phone.
Frontdesk reads the ticket, isolates the likely cause and leaves a draft reply for the team.
She asks where the release stands and hears the summary. The full answer waits in the thread.
The person on call gets an alarm in the app, and a phone call 45 seconds later.
The alert opened a thread in its room. It holds what the agent found on the server.
DeepEngine runs in a high-security bunker data center in Switzerland. Both are ISO 27001 certified: the data center, and we ourselves. The place your agents work from is as protected as the rules they work under.
In Switzerland, in a high-security bunker data center. Both the data center and we ourselves are ISO 27001 certified. Requests to the AI model are the exception: they go to the model provider configured for your organization.
A persona of DeepEngine with a name, a purpose, its own instructions, a dedicated room and its own memory. In the product it is called a bot. You connect what it should work with and decide who may use it. DeepEngine itself is the agent every workspace starts with.
Its own browser, your docs, the services you connect, and workers on your Kubernetes clusters, Linux servers and Macs. Without any of them it still answers from docs and memory.
As much as you grant. Each agent has a level: ask first, accept edits, or auto. Your restrictions apply on top and either deny an action or require an approval. Full access is not on offer.
Nothing to start. A worker is only needed on systems an agent should operate. It ships as a Helm chart for Kubernetes, a systemd service for Linux and a launchd service for macOS, and connects with a one-time install command.
Yes, in beta. Start a call in any room on the web or in the mobile app. What you say is posted as your message, the agent answers aloud, and you can give an approval by voice.
Claude or Codex. The default is set per organization, and each agent can run on the other one. Requests to the model go to its provider, Anthropic or OpenAI, with the key configured for your organization.
In a sandbox that DeepEngine operates: one isolated container per thread. A worker on your system holds no model, no provider keys and no conversation, and it only connects outbound.
Soon. DeepEngine Runtime will provide Linux and macOS runtimes that we operate for your agents. Today, agents act through their browser, through connected services and through workers on your own systems.
Book a demo and follow an agent from the first request to completed, audited work.