Session

Fireside Chat with Rohan Varma (OpenAI) and Ankit Mathur (Databricks) on Coding and Agents

Overview

Experience	In Person
Track	Artificial Intelligence & Agents
Industry	Enterprise Technology
Technologies	Databricks Agents
Skill Level	Advanced

Hear from Rohan Varma from OpenAI and Ankit Mathur from Databricks, moderated by Denny Lee (Databricks) on what it really takes to build and ship autonomous coding agents at scale. Rohan will dig into the shift from simple autocomplete to full agent development environments, and how teams can balance developer productivity with the governance, verification and controls required for enterprise use.

Session Speakers

Ankit Mathur

/Sr. Staff Software Engineer
Databricks

IMAGE COMING SOON

Denny Lee

/Director, Developer Relations
Databricks

IMAGE COMING SOON

Rohan Varma

/Product Manager
OpenAI

Full Summary

Inside the agentic shift: what OpenAI and Databricks are learning about building with Codex

A candid conversation between leaders from OpenAI and Databricks examines how coding agents are moving from autocomplete helpers to the primary interface for complex work. The discussion centers on a practical reality for builders: value now comes from delegating multi-step goals to agents, redesigning development processes around them, and tying spend to outcomes. The result is a new operating model for software and adjacent functions. Teams looking to understand how this shift lands in practice can start with what agentic analytics means for modern data workflows and how it connects to the broader platform strategy.

For teams already deploying at scale on the Databricks Platform, the message is urgent. Serving shapes and costs are changing fast, agent usage is spreading beyond engineering, and the organizations that rebuild their workflows and governance for autonomous agents will separate from those that bolt AI onto legacy practices.

Agents are escaping the code editor and expanding total output

Internal adoption of Codex at OpenAI is near universal, and non-engineers are rapidly catching up to engineers on token usage. Delegated work now includes research synthesis, outreach and follow-ups, scheduling, and spreadsheet updates, often kicked off by goal-based prompts. Code remains a powerful interface for computers, which makes coding agents effective far beyond software.

The workload is not just shifting from humans to agents, it is growing. After Codex launched as a daily driver, message volume to Codex jumped roughly 10x without a decline in ChatGPT usage, a sign that people are doing more, not the same work faster.

The economics are real, so measurement must shift to outcomes

Costs came into sharp focus. A strong P90 engineer can now spend 2,000 to 5,000 dollars per month on tokens, and the top 1 percent of users on the Ramp index reach about 90,000 dollars per user per year. Databricks rebuilt its serving stack from first principles to cope with LLM request shapes and now serves roughly 125 trillion tokens per month.

Leaders have limited time to get this under control. Limiting AI usage risks falling behind, and ignoring cost breaks unit economics. The path forward is to tie token spend to business outcomes, not input metrics. Revenue remains the clearest signal. Dashboards that reward raw consumption will encourage the wrong behavior.

From synchronous copilots to autonomous, governed agents

Most agent use still looks like a person waiting on an agent, then reviewing. The next phase is autonomous agents that run asynchronously in the cloud, verify their work, and operate under codified security and governance. The blockers are less about model intelligence and more about trust, verification, permissions, and observability.

Teams are investing in auto-review agents, security products that let leaders codify policies, and programmatic enforcement rather than repeated manual approvals. As agents touch external systems, "click yes" is not a strategy. Headless deployments for testing, incident response, and QA are emerging patterns that reduce human babysitting while increasing assurance.

Rethinking the software factory and the role of engineers

Code is no longer the bottleneck. The traditional software lifecycle, designed around scarce human coding time, breaks when agents write most of the code. Review shifts from line-by-line diffs to inspecting prompts, plans, and the human decisions that directed the agent. Teams report CI loads doubling or tripling, which forces new approaches to PR workflows and testing.

Engineers increasingly design, orchestrate, and monitor systems of agents rather than handwrite every change. Using AI to manage AI becomes essential to escape the five-to-seven-thread human limit. On the platform side, demand for smaller, faster model families and token efficiency is rising, and agent harnesses that mix models by task, such as Omnigent using Codex, help balance quality and cost.

Redesign the organization, not just the IDE

Leaders face a deployment problem, not a tooling shortage. The companies that redesign processes for secure, autonomous agents and measure success by outcomes will pull ahead. The rest will spend more for less by pasting agents onto legacy workflows.

FAQ

Agents now handle multi-hour, multi-step goals that span research, outreach, scheduling, and updates. Code is a flexible interface for controlling computers, so the same agent harness applies to finance, product, operations, and more. Internal data showed Codex usage rising sharply without displacing other tools, indicating net-new work is being done.

Input metrics like token counts drive bad incentives. The speakers emphasized tying spend to concrete outcomes, with revenue as the strongest signal. Instrument agents to connect usage to finance systems and deliverables, then manage to output, not consumption.

Model capability is often good enough. The real gaps are trust, verification, security governance, permissions, and observability. Until agents can run safely and verify their own work under codified controls, most usage remains synchronous with a human in the loop.

When agents write most of the code, value shifts to reviewing prompts, plans, and decisions that produced the change. Auto-review agents can enforce security and architectural standards, and CI patterns must adapt as agent-driven workflows increase test volume and frequency.

Tie AI spend to outcomes, not inputs. Invest in infrastructure for secure, asynchronous agents with clear permissions, observability, and automated reviews. Ensure senior leaders actively redesign workflows and governance, since process change, not smarter models, is the short-term unlock.