Introducing Project Agens

Intelligence,
made for work.

Advancing agentic foundation models and systems
and the next generation of AI-native work.

Project Agens

Project Agens is an MSR initiative building an end-to-end-optimized foundation model and system purpose-built for long-horizon agents and the Microsoft ecosystem. It is designed to be agent-first, frontier-efficient, multimodal-native, with built-in multiagent and continuous learning capability, transforming the fundamental and frontier AI research and innovation at MSR into Microsoft advantage on AI models and systems. The initial MVP aims to build frontier work intelligence and work agents for AI-native work experiences. The first milestone is a 300B MoE model targeting 10–100× cost efficiency and 80–90% traffic replacement against frontier proprietary and open-source models, starting with work and enterprise scenarios.

SELECT A LAYER
THE MODEL

Agens

  • Agent-first
  • Frontier-efficient
  • Multimodal-native
  • Multiagent-native
  • Continuous learning
First milestone300B MoE

Purpose-built for the Microsoft ecosystem, starting with work and enterprise scenarios.

Explore the Agens Model

The model, harness, and optimization pipeline can extend to other products.

Next: Agens 2.0
The Agens model

Frontier efficiency.
Purpose-built intelligence.

Explore Agens Model

An end-to-end-optimized agentic foundation model, starting with work and enterprise scenarios across Microsoft products.

300B

MoE Agens model

10–100×

Cost-efficiency target

Against frontier proprietary and open-source models.

80–90%

Traffic-replacement target

Across Microsoft products, starting with work and enterprise.

About the first-milestone targets+

The first milestone is to build a 300B MoE Agens model, targeting 10–100× cost efficiency and 80–90% traffic replacement against frontier proprietary and open-source models across Microsoft products, starting with work and enterprise scenarios. These figures describe the project’s first-milestone targets.

Agensh · Agens HarnessResearch · September 2026

Scalable agents.
Self-organized intelligence.

Scaled to 1,024 agents

Demonstrated on ProgramBench’s pandoc task. Workers claim tasks, coordinate with peers, and merge verified progress asynchronously—without a central orchestrator.

Equal-sized agent nodes form distributed peer networks, with connections within and between groups and no central orchestrator.
Concurrent workers. Asynchronous progress.Conceptual illustration
Agens Work04 / EXPERIENCE

A new way to work.

Explore Agens Work

Next-generation AI-native work experience, grounded in the Office and Microsoft 365 ecosystem.

Your proactive, always-on work agent.

Illustrative workflow

Find work that
needs doing.

Connect decisions, customer priorities, and work in progress to find what needs attention—before you ask.

Teams

Rollout delayed by two weeks

Outlook

Data residency becomes a priority

Word

Customer review last updated 12 days ago

Work detectedThe customer review needs an update.

Shared live context

Where work happensTeams · Outlook · Word · PowerPoint · Excel · Microsoft 365

Continuous learning AI (aka Agens 2.0)

The next horizon

Intelligence that keeps learning.

Build the continuous learning loop and infrastructure to empower every organization with sovereign intelligence

ExperienceConsolidateLearn
01Online learning+

A new paradigm of learning: intelligence that learns how to learn as new challenges emerge.

TEST-TIME LEARNING: TRAINING AND DISTILLATION
02Continual learning+

Continuous reinforcement learning, with the ambition to build new capabilities without forgetting what came before.

CONTINUOUS RL + LEARNING WITHOUT FORGETTING
03Experiential learning+

Learn from experience. Consolidate knowledge. Coach and improve through the work itself.

EXPERIENCE → CONSOLIDATION → COACHING
Project Agens

Let’s move work forward.

Explore research

Search Project Agens

Esc