Digital twin
A digital twin, its wiki-LLM memory, and WhatsApp as the command center
Every person (and the company itself) gets a digital twin. Its memory takes the form of a wiki that the model rewrites over time, following Andrej Karpathy's "LLM as an operating system" idea. Linked to its owner's WhatsApp, the twin becomes their command center: you text it, it orchestrates the agents and comes back with the result.
Under the hood- WhatsApp paired with each team member's number: the twin receives requests there, delegates them to agents, asks for approvals and delivers files in the conversation.
- Memory structured as knowledge pages: people, accounts, projects, decisions.
- The model re-reads, corrects and merges its own notes: memory stays short and current.
- Targeted retrieval: for each task, only the relevant pages enter the context.
In plain termsYour agents know your context from the first request, without you re-explaining everything. And everyone runs their team of agents from WhatsApp, no app to open.
Execution memory
Each agent has its own working memory
Beyond the shared company memory, each agent keeps an execution memory: what it tried, what worked, the preferences it learned on its recurring tasks.
Under the hood- Episodic trace per run: goal, steps, tools called, result and human approval.
- Successful patterns become habits; failures become guardrails.
- Isolated per agent: sales doesn't inherit support's reflexes, and vice versa.
In plain termsThe agent improves with use: its follow-ups, quotes and support replies get faster and more accurate.
Nightly dreams
At night, the agents prepare the next day
Outside office hours, the agents replay the day's tasks, explore leads and prepare drafts.
Under the hood- Replay of past tasks to extract reusable lessons.
- Forward exploration: drafts, monitoring, hypotheses prepared for the next day.
- Asynchronous work: heavy processing runs outside office hours.
In plain termsIn the morning, you find leads, drafts and the points that need your attention.
Consolidation
Nightly consolidation of the company brain
Just as a brain consolidates memories during sleep, Bruno consolidates your company brain every night: the day's scattered traces are sorted, linked and integrated into long-term memory.
Under the hood- Agents' execution memories are summarised and promoted to the company wiki.
- Deduplication and linking: scattered facts become connected knowledge.
- Useful information is reinforced and the rest is pruned, so memory stays compact.
In plain termsWhat your company knows stays available, even when someone leaves or information gets lost in a thread.
Orchestration
The agents talk to each other
The agents share out the work. Sales asks for research, the writer pulls the data, the twin arbitrates. You approve the key steps.
Under the hood- Internal exchange protocol: an agent can delegate a sub-task to another.
- Orchestration by the twin: it routes the request to the right roles and recomposes the result.
- Human checkpoints: you approve the steps that commit the company.
In plain termsYou set a goal and the agent team organises itself to deliver it.
Computer
A persistent Linux environment for agents to execute their work
Computer gives Bruno an isolated Amazon Lightsail server to run code, install tools, process files and serve an application. It is offered across all business files; each server remains scoped to its business file and owner.
Under the hood- Ubuntu 24.04 machines on Amazon Lightsail in three resource classes. Bruno’s AWS role handles provisioning; SSH connections use temporary certificates and the server’s private address, with host key verification.
- Persistent processes, working files, snapshots, restoration and work branches. Completed code projects can be backed up to a private GitHub repository when that backup is configured.
- Outbound networking is blocked by default, then opened for explicitly authorised domains. Business data reads use temporary grants limited by duration and volume; Bruno’s brokers retain source credentials.
- Applications are served through a Bruno HTTPS gateway, with revocable links and a selected audience. Publishing an application does not grant access to the server’s terminal or files.
- Server creation, deletion, network changes and data access require the applicable approvals. Actions and usage are recorded; mission completion is recorded after the result is verified.
- A maximum monthly hosting price is shown before creation, separately from model and agent work costs. The server stays active and billed until explicitly deleted. A graphical desktop is not available on this runtime.
In plain termsYour agents have an environment that retains files and applications between conversations. You can resume the work, retrieve results and choose who can view the application.
GPT-4o€0.012
Claude€0.008
Mistral€0.002