microsoft/Sico

microsoft/Sico

Description: An open-source Digital Worker platform for reliable execution and continuous co-evolution.

Language: Go

License: MIT

Stars: 31

Forks: 2

Open issues: 4

Created: 2026-03-09T13:44:18Z

Pushed: 2026-06-12T06:05:17Z

Default branch: main

Fork: no

Archived: no

README:

What is Sico?

Sico — Symbiotic Intelligence for CO-evolution — is an open-source platform for building, managing, and evolving Digital Workers: structured AI labor units that co-evolve with human operators through real production work, particularly in BPO (Business Process Outsourcing) scenarios.

The idea behind Sico emerged from large-scale operational challenges observed in Microsoft’s internal environments, especially across BPO-style workflows such as black-box testing.

Through real production workloads, Sico achieved closed-loop validation for Digital Workers operating under continuous execution, evaluation, and human supervision. Through this process, we observed that reliability emerged not from static automation alone, but from the continuous co-evolution between human operators and Digital Workers.

In Sico, four core roles define how work gets done:

• Employer: defines business objectives and outcome standards for Digital Workers
• Operator: responsible for training, monitoring, and improving Digital Workers
• Developer: builds the capability structure and infrastructure (workflows, tools, execution environments) of Digital Workers
• Digital Worker: executes tasks through structured capabilities and continuous learning

At the center of this system, a Digital Worker is not just a model or an agent, but a structured, executable capability unit.

Its anatomy consists of:

• Cortex: reasoning and planning (LLMs, agent frameworks)
• Action: execution via domain skills, workflows, and sandboxed tools
• Memory & Sense: accumulated knowledge, execution experience and contextual awareness for grounding and continuous improvement

This creates a practical Co-Evolution loop where humans and Digital Workers continuously improve together through real work.

For a comprehensive survey of this direction, refer to [Agentic Evolution: From Self-Improving Agents to Co-Evolving Human–AI Systems ](docs/agentic-evolution.pdf)

> Learn more: [What is Sico](docs/overview.md)

Who is Sico for?

Sico is primarily designed for:

• BPO providers looking to build and operate AI-powered workforces for their clients
• Enterprises seeking to automate and scale operational workflows with AI
• Developers building AI workers for specific business domains

Why Sico?

Many real-world workflows, especially in BPO scenarios such as black-box testing, data processing, customer support, and content moderation, require continuous, stable execution at scale.

BPO is a natural environment for Digital Workers:

• structured evaluation signals are continuously produced • feedback loops naturally exist • execution and supervision responsibilities can be clearly separated

Traditional automation approaches rely on static scripts or predefined workflows. However, production environments continuously change:

• interfaces evolve
• rules shift
• data formats change
• edge cases emerge

As a result, automation often becomes brittle and requires repeated manual adjustment.

Digital Workers approach this problem differently. Instead of treating execution as a fixed process, Sico treats execution as an evolving capability.

As Digital Workers take on execution, human roles shift from doing tasks to guiding evolution through the Operator role.

Each completed task contributes signals that help Digital Workers adapt to real environments, enabling organizations to scale execution capacity while continuously increasing reliability.

| Pain point | Sico's approach | | --- | --- | | Agents are thin wrappers around a model and a toolbox | A structured Cortex / Action / Memory architecture with project-level knowledge | | AI repeats the same mistakes task after task | Execution experience captured as training signals for continuous improvement | | Full autonomy is unreliable; humans can't easily intervene | The Operator role: human-in-the-loop collaboration with clear responsibility boundaries | | GUI automation is flaky and hard to reproduce | Sandbox execution with isolated environments, step-level traces, and replayable runs |

Architecture at a glance


Frontend (React) ──HTTP/SSE──▶ Nginx ──▶ Backend (Go / Gin)
│ ▲
gRPC │ │ reverse gRPC
▼ │
Core (Python / asyncio)

• Backend: HTTP APIs, persistence, RBAC, sandbox orchestration.
• Core: agent reasoning, tool execution, LLM orchestration, experience accumulation.
• Reverse gRPC: Core calls back into the Backend to persist messages, update state, and send notifications, so the Core remains database-free.

On top of this runtime, Sico organizes work into three loops that together form the co-evolution cycle between Operators and Digital Workers:

• Execution Loop: turns an Operator goal into a traced agent run. The Cortex–Action–Memory stack executes inside an observable Sandbox and emits structured trajectories: actions, intermediate states, tool outputs, and environmental feedback.
• Evolution Loop: converts those trajectories into reusable capability. A Reflector → Curator pipeline distills successful strategies and recurring failure patterns into a per-(project, agent) Playbook that is injected into the next run's workspace (training-free), while the same signals can also be fed back into base-model training (training-based).
• Evaluation Loop *(planned)*: analyzes failed task trajectories and attributes the root cause using an L1–L4 failure taxonomy, from high-level ownership to concrete failure mode. The results help the Operator provide targeted corrections and feed failure insights back into Experience Learning and future training.

Features

• Multi-service platform: Go backend, Python core, React frontend, communicating over gRPC and a bidirectional *reverse gRPC* pattern.
• LLM Hub: unified model runtime with adapters for OpenAI, Azure OpenAI, Anthropic, Gemini, OpenRouter, OpenAI-compatible providers, and generic HTTP JSON / binary endpoints. See [LLM Hub docs](backend/docs/llmhub.md).
• Agent toolkit: file I/O, grep, shell/command execution, web search &...