Microsoft has introduced a preview of the Microsoft Agent Framework, an open-source improvement package for .NET and Python for creating AI brokers and multi-agent workflows.
It helps creating particular person brokers in addition to graph-based workflows to attach up a number of brokers.
In line with Microsoft, the Agent Framework is a direct successor to its different initiatives Semantic Kernel and AutoGen, using foundations from each. It brings collectively Semantic Kernel’s enterprise-grade options like thread-based state administration, kind security, filters, telemetry, and mannequin and embedding assist, with AutoGen’s abstractions for single- and multi-agent patterns.
“Each had passionate customers — however every had gaps,” Microsoft wrote in a weblog submit. “Builders requested us: why can’t we’ve each — the innovation of AutoGen and the belief and stability of Semantic Kernel — in a single unified framework?”
Like Semantic Kernel and AutoGen, the Agent Framework may even be open-source and Microsoft welcomes the group to contribute.
The Agent Framework was constructed round open requirements, and helps MCP, Agent-to-Agent (A2A), and OpenAPI. MCP permits brokers to find and use exterior instruments or information sources, A2A permits builders to construct brokers that may collaborate with different brokers, and OpenAPI assist permits any REST API that follows the OpenAPI specs to be imported as a callable software. Brokers can run in containers, on-premises, or throughout a number of clouds.
It helps a number of totally different orchestration patterns, together with sequential orchestration for step-by-step workflows, concurrent orchestration, group chat orchestration for brokers to collaborate, handoff orchestration to maneuver duties between brokers, and magnetic orchestration, which permits supervisor brokers to construct a process ledger that coordinates specialised brokers and people to unravel complicated issues.
Different options that the Agent Framework has contains built-in connectors to enterprise methods, pluggable reminiscence modules, YAML and JSON definitions for declarative prompting, assist for OpenTelemetry, and the power for brokers to pause, resume, and get better from interruptions.
“Brokers are quick changing into the subsequent layer of software logic — reasoning about targets, calling instruments, collaborating with one another, and adapting dynamically. With Microsoft Agent Framework, builders now have a single, open-source basis that carries the perfect of analysis innovation into manufacturing with the sturdiness, observability, and enterprise readiness required to scale,” Microsoft wrote.