Open-source research and engineering for multimodal AI, autonomous agents, and AI-native workflows.

Frontier AI is a research unit within Sber AI focused on developing advanced artificial-intelligence technologies with a long-term orientation toward stronger, more autonomous AI systems, including progress toward AGI. The team studies and adapts state-of-the-art AI methods while building its own infrastructure for agentic research, experimentation, code generation, forecasting, memory, observability, and automation.

Our work is centered on three connected directions:

• Multimodal models — systems that can understand and reason across text, images, audio, video, and other information streams.
• Fundamental AI research — new algorithms, architectures, evaluation methods, and workflows for processing and generating information at increasingly human-like levels of capability.
• AI agents — autonomous systems that do more than answer prompts: they form hypotheses, write code, call tools, run experiments, analyze evidence, and improve workflows.

The repositories in this organization represent the open-source part of that work: graph-based multi-agent runtimes, workflow studios, deep-agent control planes, framework-neutral memory, forecasting workspaces, and DevOps agents. Together, these projects support the transition from classic software automation toward AI-native work: systems where agents can plan, execute, observe, learn, and collaborate with humans and tools.

• LinkedIn: TODO — add organization page URL after it is created or confirmed

This organization currently contains repositories focused on:

• Agentic research infrastructure — explicit graph runtimes, agent workspaces, and harnesses for hypothesis-driven experimentation.
• Multi-agent systems — LLM agent teams with roles, tools, budgets, memory, topology changes, and observable execution.
• Agent observability — run timelines, tool calls, token usage, events, metrics, diagnostics, topology changes, and error traces.
• Deep-agent orchestration — phase-driven control planes, watcher processes, evidence-based selection, and human-in-the-loop checkpoints.
• Agent memory — typed long-term memory blocks, durable backends, lifecycle policies, graph links, promotion gates, and evaluation harnesses.
• Forecasting and evidence analysis — evidence collection, adversarial analyst debate, calibrated probabilities, source provenance, reports, and audit bundles.
• AI-native DevOps automation — chat-driven operational agents for shell, deploy, GitLab, Docker, AWS EC2, and DNS workflows.

Dynamic graph runtime for building, adapting, and observing multi-agent LLM systems in Python.

• Documentation: frontier-ai-next.github.io/gMAS
• Quick start: gMAS quick start
• Language: Python

gMAS is a Python framework for graph-based multi-agent systems. Nodes hold roles and tasks, edges carry work and context, and the runner schedules model and tool calls while keeping budgets, memory, events, outputs, latency, errors, and topology changes visible in one execution model.

Graph-native studio, runtime integration, and self-hosted observability for mutable multi-agent LLM workflows.

• Live demo: gmas.frontierai.ru
• Documentation: frontier-ai-next.github.io/flowMAS
• Core SDK: gMAS
• Language: TypeScript

flowMAS provides a workflow studio, demo API, and observability layer around gMAS: visual graph editing, execution-order preview, LLM provider and tool configuration, run monitoring, run history, metrics, diagnostics, and error output.

Control panel for deep-agent harnesses and phase-driven agent execution.

• Live demo: harness.frontierai.ru
• Start here: docs/vision.md
• Language: Python

Umbrella is a workspace-first, phase-driven control plane for interchangeable deep-agent heads. It can run several agent heads per phase in harness mode, select winners with evidence, guard runs with watcher processes, and keep humans in the loop through checkpoints and critique restarts.

Framework-neutral memory layer for agent systems.

• Language: Python

Lethyx packages typed agent memory blocks, durable backend adapters, lifecycle policies, graph links, promotion gates, and evaluation harnesses behind a small public API. It is designed to add useful long-term memory to frameworks such as LangGraph, MCP servers, REST services, and plain Python agents without making memory a second control plane.

Live gMAS workspace for forecasting user-selected world events with parallel evidence collection, adversarial analyst debate, and continuous follow mode.

• Live demo: forecast.frontierai.ru
• Language: Python

forecast-ai rewrites user prompts into canonical forecasting questions, creates typed forecast specs, routes different horizons through deterministic policies, collects public signals, runs analyst debate, produces calibrated outcomes, and stores forecasts, source records, reports, audit bundles, and run history.

Skill-guided DevOps agent built on MECE.

• Live demo: mgarlbot.frontierai.ru
• Paper and benchmark: submission/README.md
• Language: Python

mgarlbot is a DevOps-focused agent for chat-driven operations across shell, VPS deployment over SSH, Docker, GitLab API, AWS EC2, and REG.RU DNS.

• Use gMAS if you need a Python runtime for explicit graph-based multi-agent systems.
• Use flowMAS if you want to design, run, and observe gMAS workflows through a web studio.
• Use Umbrella if you are exploring phase-driven deep-agent orchestration and harness evaluation.
• Use Lethyx if your agent needs durable, typed, evidence-aware long-term memory.
• Use forecast-ai if you want an applied gMAS forecasting workspace.
• Use mgarlbot if you are interested in agentic DevOps automation.