When Agents Negotiate With Agents with Microsoft Research's Saleema Amershi

10 snips

Feb 28, 2026

Saleema Amershi, Partner Research Manager at Microsoft Research who leads teams behind AutoGen and Magentic-One, discusses multi-agent AI systems. She explores agent-to-agent commerce, the Magentic Marketplace, risks like manipulation and first-proposal bias, and tradeoffs between walled gardens and open ecosystems. She also considers human oversight, duty of care, and how interaction guidelines must evolve.

Ask episode

AI Snips

Chapters

Transcript

Episode notes

INSIGHT

Agents Shift From Orchestration To Complex Collaboration

Agents have moved from isolated chat assistants to entities that can act in the world, shifting problems from orchestration to collaboration.
This raises cross-cutting challenges like privacy, security, and reasoning about other agents when full information can't be shared.

ANECDOTE

Magentic One Demonstrated Orchestrator And Skill Agents

Magentic One used an orchestrator plus specialized skill agents (web surfer, file surfer) to complete complex web tasks.
Magentic UI was built as a research platform to study human oversight with features like co-planning and intervention prompts.

ADVICE

Keep Humans In The Loop With Co‑Planning And Checkpoints

Keep humans in the loop for long-running agent tasks via co-planning and checkpoints.
Present upfront plans and surface irreversible or risky actions so users can intervene before errors cascade.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

This episode is sponsored by Airia. Get started today at ⁠⁠⁠⁠⁠⁠⁠⁠⁠airia.com⁠⁠⁠⁠⁠⁠⁠⁠⁠.

Saleema Amershi is a Partner Research Manager at Microsoft Research AI Frontiers, where she leads the teams behind AutoGen and Magentic-One. She joins the AI Inside podcast to discuss what happens when AI agents stop working in isolation and start collaborating, transacting, and negotiating with each other at scale. We cover the challenges of multi-agent orchestration, why current frontier models show a dangerous first-proposal bias that favors speed over quality, and how her team built the Magentic Marketplace to simulate agent-to-agent commerce. We also dig into agent manipulation, the tension between walled gardens and open ecosystems, and what duty of care means when an AI is acting on your behalf.

Note: Time codes subject to change depending on dynamic ad insertion by the distributor.

Chapters:

0:00 - Podcast Start
01:46 - Introduce Saleema Amershi
0:02:04 - Talk about the challenges surrounding the orchestration of these agents together. And what limits this collaboration today, is it a technical limitation or something deeper?
0:05:10 - OpenClaw perspectives
0:06:42 - From your research, why would these projects fall apart? And actually, are companies setting themselves up for failure by moving too fast?
0:07:53 - What IS Magentic AI
0:12:16 - Collaboration between humans, AI and the agents.
0:18:45 - Agents eliminate the scarcity of attention
0:20:04 - You found that these frontier models that we're often told are ready to handle complex tasks like these were getting stunted by too many choices. They were showing a strong bias toward whatever option showed up first regardless of the quality. So, what's going on there?
0:24:07 - The first-proposal bias
0:33:25 - You co-published 18 guidelines for human-AI interaction back in 2019. When you look at those original guidelines through the lens of what you're working on today, what still holds up and what could you not have anticipated back then?
0:36:58 - What are the lessons for marketers and sellers?

Learn more about your ad choices. Visit megaphone.fm/adchoices