Anton and working channels
Goals, initial context, critical decisions and confirmation of actions with external impact.
Live project · Architecture · Operations
A digital twin
that learns to work
Digi Anton is an evolving multi-agent system for research, knowledge, content, engineering and operational tasks. It combines local compute, long-term memory, specialised roles, recurring cycles and human control.
4linked compute nodes
8specialised roles in a single system
Dailyoperational reports and checks
Initial task
Not another chat.
A working digital system.
Personal and project work is spread across documents, tasks, research, code, visual materials and infrastructure. A single universal assistant quickly loses context and cannot show what was actually done.
Digi Anton's task is to separate responsibility between roles, persist decisions in long-term memory, execute repeatable work and present an observable system state. Autonomy grows incrementally, only after a specific system is validated.
Architecture
Orchestration separate.
Heavy compute separate.
Orchestration platform on a dedicated control node
A universal tool harness, task routing, queues, roles, statuses, reports and external service integration through a single managed system.
Specialised roles
Content, knowledge, business analysis, visuals, compute, programming and personal support.
QMD and local RAG
Documents, accepted decisions and search with source citations.
Specialised compute nodes
GPU system for visual and applied models, and a separate two-node system for heavy local analysis. Models are chosen by quality, speed and task cost.
Separation of responsibility
Eight roles.
Not eight identical chats.
Orchestrator
Decomposes the goal, selects the route and assembles the result.
Knowledge manager
Maintains memory, sources and accepted decisions.
Content planner
Research, plans, articles and publish packages.
Business analyst
Structures tasks, data, options and conclusions.
Visual agent
Images, visual briefs and source control.
Compute operator
Routes resource-intensive work to the compute node.
Dev team
Code, tests and changes with engineering safety gates.
Personal secretary
Task reviews and preparation of actions requiring verification.
Memory
Decisions must not disappear with the chat
Documents, project files, reports and canonical rules.
Long-term knowledge base, usable by humans and agents.
Email, calendar, tasks, Telegram, browser and other services connect as swappable tools, not baked into a single model.
Read-only retrieval of relevant context with source tracing.
Accepted decisions are written before work continues and treated as canonical.
Operations
Autonomy must be observable
Daily report
Status of roles, queues, errors, models, memory and backup cycles.
Health-checks
Checks of key service availability and reproducible working routes.
Redundancy
Local copies and mutual replication of core systems across nodes.
Risky actions
Publishing, external changes and sensitive operations require approval.
Practical takeaways
What we've already learned
A useful process matters more than new infrastructure.Start with a real repeatable task, then automate.
Roles require different contexts and permissions.An agent's name alone doesn't create specialisation.
Memory needs rules.Without decision canonicalisation, RAG confidently returns stale options.
Autonomy without reports is invisible.Observability is needed for both trust and error correction.
Local models don't replace quality control.Route is chosen by task, risk, cost and required outcome level.
Honest boundaries
What we do not claim is ready
Independent external publishing, message sending, calendar writing, voice channels and financial decisions without human involvement are not claimed as fully autonomous capabilities. The system evolves iteratively; status of each system matters more than a flashy demo.
Application
Let's start with one repeatable task
Describe the process, sources, frequency, cost of error and the result you currently assemble manually.
admin@akonnov.ru