Ishir Garg

student at @berkeley_ai

[Ishir Garg @ishirgarg]

@m_emanaftab There’s an implicit assumption that the offline data should be sufficiently diverse. However our method can “stitch” across trajectories. For example we recover the empowerment on antsoccer-stitch where the ant either dribbles or navigates with no ball, but never combines both

[Ishir Garg @ishirgarg]

How should an agent decide where to explore, given unlabeled offline data? 🤖

In our new paper, we show that empowerment, a measure of how much control an agent has at a state, can be estimated from offline data and used to guide exploration!

https://t.co/O5tz9nQTsB
(1/7)

[Ishir Garg @ishirgarg]

@pranav_modi @vivek_myers @qiyang_li @ben_eysenbach @svlevine 😅 fixed!

[Ishir Garg @ishirgarg]

🌐 Project page: https://t.co/O5tz9nQTsB
📜 Paper: https://t.co/x4ilWXODDf
🧑‍💻 Code: https://t.co/2nzOpBn6z2

Thanks to my amazing collaborators @vivek_myers, @qiyang_li, Catherine Ji, @ben_eysenbach, @svlevine!
(7/7)

[Ishir Garg @ishirgarg]

How do we use this for online goal-conditioned RL?

Rewarding the agent for reaching high-empowerment states helps it solve sparse-reward, goal-reaching tasks faster!
(6/7)

[Probe @useprobeai]

Introducing Probe: The Agent-Native Research Stack.

Our goal is simple: help AI/ML teams run better experiments, faster.

Teams run research with agents now, but their work is disordered. They lose hours verifying what their agents actually did, managing environments and compute, and re-explaining company context their agents should already know.

So we rebuilt the research stack from the ground up for agent-driven research.

Probe brings core research tools, shared context from past experiments, and visibility into each run together in one platform.

Teams can spend less time setting up infrastructure, supervising agents, and catching teammates up, and more time testing ideas.

If you’re a research lab, scale your research team without sacrificing speed.

If you’re an AI/ML team, stand up a research capability with the infrastructure of a frontier lab.

See Probe for yourself: https://t.co/FXWNS9wLzU

[Narada AI @Narada_AI]

Narada is now #1 on the official WebArena leaderboard. Narada Code + Operator achieved an 86.1% task success rate, 11.8 percentage points above the previous best reported result.

WebArena measures whether agents can complete multi-step tasks across realistic web applications. It addresses a challenge enterprise automation teams face every day: getting an entire workflow done correctly.

Leading an automation CoE? Bring us a repeatable workflow your team still handles manually. Whether it involves moving data between systems, updating records, or navigating multiple applications, let’s explore how Narada could help meet your requirements for accuracy and reliability.

📅 Book a demo to discuss your workflow: https://t.co/Yqg7NLRI7T

#RPA #IntelligentAutomation #AgenticAutomation #ProcessAutomation #AI #Agents #AIAgents

[Ishir Garg @ishirgarg]

Our paper, MemFail, is accepted to #EMNLP2026!

What happens when memory systems fail? We introduce MemFail, a framework for analyzing their failure modes...the results may surprise you!

📄: https://t.co/zOjCIvBpKH

Thanks to my collaborators Neel, @dawnsongtweets, @xuandongzhao

[Narada AI @Narada_AI]

What does it take for an AI agent to complete a complex enterprise workflow reliably, not just once, but at production scale?

At UC Berkeley’s 2026 Agentic AI Summit, Narada AI CEO and co-founder Dave Park discussed why so much enterprise work remains manual:
- APIs often cover only part of a workflow
- Hard-coded scripts and RPA break when real-world conditions change
- General-purpose AI agents can lose context during processes spanning hundreds of steps

Dave shared how Narada addresses these challenges by combining planning, validation, guardrails, exception handling, and self-healing with UI automation that works across web applications, APIs, Citrix, VDI, remote browsers, and legacy systems.

We’re also proud to have been selected for the Berkeley Xcelerator’s 2026 Spring Cohort.

A special thank you to @BerkeleyRDI for organizing the Summit and bringing together an inspiring community of researchers, founders, and builders advancing agentic AI.

Read the recap: https://t.co/26W1wI7ZQU
Watch the full talk: https://t.co/O9HVeD2B8X

#AgenticAI #EnterpriseAI #Automation #AIAgents #UCBerkeley

[Ishir Garg @ishirgarg]

@xuandongzhao @berkeley_ai @OpenAI Congrats!!