Saejong Lee retweetou

Microsoft introduced an Agent framework this week, and it went completely unnoticed.
The agent team can browse the internet, internal files, execute code, and more.
In this example, the Microsoft team made the agent autonomously order a sandwich.
First, an agent:
A) Plans - breaking a large complex into smaller tasks
B) Acts - including executing code, using more efficient tools
C) Observes - processes data to determine the next step
D) Reflects - determines whether it was successful in the task
The framework, Magentic-One, attempts to co-ordinate multiple agents working together on a task
It formalizes this process by making a language model “Orchestrator” asks a series of five questions after each step taken by individual agents:
1) Is the request fully satisfied (i.e., task complete)?
2) Is the team looping or repeating itself?
3) Is forward progress being made?
4) Which agent should speak next?
5) What instruction or question should be asked of this team member?
- The Orchestrator also manages timeouts and failures, with an outer control loop, and an inner control loop
-MSFT created specialist agents: a Websurfer, Coder, FileSurfer, ComputerTerminal user to work together on tasks
-They find that their framework is competitive with (but not beating) State of the Art on agent benchmarks such AssistantBench and GAIA
MSFT’s work is interesting as you can visualize a future where firms can provide hyper-specialized agents that are very good at a particular task and can work together with other agents to solve larger problems over time.
And maybe we can finally clear the JIRA ticket backlog…
Hussein Mozannar@HsseinMzannar
Excited to release our agent team Magentic-One! Magentic-One can browse the web, files, write and execute code & supports a human-in-the-loop. Built by @MSFTResearch AI Frontiers with @pyautogen. aka.ms/magentic-one-b… My favorite task:
English


























