Aug 28, 2026 |

6 min

|

1k views

Calibrating Autonomy: Why the Best AI Agents Know When to Step Back | CDO Vision Los Angeles 2026

Chetan Thapar, who leads product for Posit's Snowflake integration, shares his take on agent autonomy and what Posit is building next.

What's covered:

- Why coding assistants, analytical agents, and AI-powered support triage are "table stakes," and where the real value building happens next
- Posit's three areas of focus for 2026: a unified AI agent spanning the full data-to-decision pipeline, dynamic visualization powered by semantic models (no more pre-canned dashboards), and deterministic skills that pair LLM orchestration with statistically validated machine learning methods
- "Calibrating autonomy": why data science workflows need agents to run independently on verifiable tasks, but stay human-in-the-loop for exploratory analysis, where the real insights often come from noticing something strange in the data
- The difference between a plausible AI answer and an actually correct one
- Why the next shift in AI is moving from systems that answer questions well to systems that help people ask better ones (with a nod to how Claude Code prompts the prompter)

Learn more about Posit for Snowflake: https://posit.co/solutions/snowflake

This interview is from CDO Vision Los Angeles March, 2026 (AIM Media House's invite-only summit for Chief Data Officers and enterprise AI leaders)

Transcript#

This transcript was generated automatically and may contain errors.

Hi everyone, I'm Chetan Thapar and I lead product for snowflake integration at Posit, which is the code-first open-source data science company, and I'm really excited to be here at CDO Vision Los Angeles. It's nice weather and good people and good conversations.

Interesting question because AI just moves so fast, but if I look internally, our engineering team is using coding assistants to basically build all the code flow to do reviews and just accelerate supercharged productivity. We have our data science team building agents on analytical pipelines. We have AI being used for sales workflows, for support, triage and so on, but honestly that all is table stakes. What we are spending, what I'm spending most of the time is building value for customers.

And so if you think about Posit, we have a full pipeline, so it goes all the way from data to decision. And if you look at the pipeline, that is mostly coding, thinking and reviewing. And those are the things which are really accelerated by AI. So I'm really excited about that and we have delivered a foundation already, which is building agents for exploratory analysis, for coding and debugging, as well as self-service analytics.

Three focus areas for 2026

What we are moving to next is three areas. So we're building a unified AI agent, which extends all the way from data to decision. So a data analyst in one sitting can go from raw data to a self-service natural language powered analytics dashboard using grammar of graphics, all in one sitting, but powered by respecting semantic borders. So I think that's number one, the unified AI, the learners get semantic borders.

The second is ensuring that end users can interact with the data, again, powered by the same semantic borders, but using dynamic visualizations. So we don't like pre-canned dashboards. I think that's gone. What we need is an end user asks a question and using grammar of graphics, they can get rich visualizations, which are reproducible, auditable and validated. So I think that's the number two area that we are really focusing on.

And the third thing is deterministic skills. So think of AI workflow in which LLM still orchestrate and agents still orchestrate, but you have statistical validated machine learning methods, which are called as skills. So those are the key three areas which we are really focusing on and which I think are new in 2026.

Calibrating autonomy

It's essentially in two words, calibrating autonomy. So there is a prevailing notion or the prevailing narrative that more autonomy in AI agents is always good. And human in the loop is friction or a stopgap. We believe that data science workflows are slightly different. And so we don't fully agree with that. We think that there are different jobs, different tasks to be done in a data science workflow. And each one of them has to be calibrated on the spectrum of autonomy.

So something which is very much verifiable, like building a quick dashboard, human is our friction in that. Let the agent run his job, do his job and complete it. But there are certain things like exploratory analysis in which if the agent just gives you the answer, you actually miss out on the insights that really matter. Because data scientists, when they are doing exploratory analysis, they find something strange, maybe a weird correlation or some kind of statistical anomaly. And that allows them to double click on something and find novel insights.

And so the hardest problem is finding a design pattern which can differentiate between these two ends of the spectrum. And so our approach is if something is verifiable, let the agent run. If it requires exploratory analysis or very consequential, in that case, the agent shows insights and then we expect the data scientist to take the seat. So lean in versus lean back. And I think that's kind of key for us.

So lean in versus lean back. And I think that's kind of key for us.

And it's because all of the products that we build are open source and code first, we can enable data scientists to do that. So again, calibrating autonomy, I think, is the hardest problem for agents right now.

From plausible answers to the right questions

I think it's a gap between what a plausible AI answer is, a plausibly correct AI answer is, and what actually correct AI answer is. And I think, again, going back to the calibration of autonomy, many data AI agents right now are optimized for, let's call it, obedient autonomy. You give it a question and it just gives you a polished answer and that's it. But that polished answer is sometimes not the correct answer.

The trend to watch out for is moving from an AI which answers questions really well to an AI which allows people to ask the right questions. And we can see that actually just outside our sector as well. So if you look at Anthropic's cloud code, it kind of prompts the prompter. So when you ask a question, it asks you more clarifying questions rather than just assuming intent. And so that, in which we provide a data scientist with context and insights and guide them to a certain direction and let them operate at a higher level of asking the right questions, is what we have built in our product. And I think that's the trend we need to be really looking forward to in the next year.

The trend to watch out for is moving from an AI which answers questions really well to an AI which allows people to ask the right questions.

I mean, there's no dearth of AI content. There's webinars , there is blogs and a lot of that. But I think what CDO Los Angeles or CDO in general gets right is you all are getting practitioners all together. And they are talking not just about what's theoretically possible, but what exactly is the bottleneck in data science workflows, in execution of AI and adoption of AI. And I think that is critical. And for me personally, all the leaders that are here are the industries for which we build products. So healthcare, financial services, insurance and so on. So what I'm looking for is really finding out in practice what's working for them and what's not working for them. Because that feedback loop allows us to just build the products better and just make the industry go forward in general. So really excited about that.