DevRev has developed two agents, named Computer and Operator, to assist service and operations teams in completing complex customer and internal requests. The company says the agents are designed to reason over live company data, that actions are reversible, and that responses link back to their sources, according to Ahmed Bashir, DevRev’s chief technology officer.
Models, integration and goals
DevRev integrates OpenAI models—including GPT‑6 Astra and GPT‑6.1 Sol—into a platform that connects corporate knowledge and gives teams control over what actions an agent may take. The firm positions OpenAI’s models as providing the reasoning and reliability required for these tasks while meeting practical constraints around latency and serving cost.
Ahmed Bashir summarized the trade-offs: "It’s a combination of being able to have the right reasoning capabilities, being able to be trusted, highly aligned, and most importantly, it has to work from a latency perspective."
How the system is built and evaluated
OpenAI models drive the reasoning, planning, and tool selection for both Computer and Operator. In DevRev’s Agent Studio, users can specify how they want a tool to behave in conversation and lock parts of its behavior behind fixed rules. DevRev tests the models together with the software that supplies company context, enforces permissions, and controls actions.
The company uses an internal benchmark called Enterprise‑Bench to assess performance in realistic conditions where data exists in multiple versions, is stored across different systems, and access depends on a person’s role.
According to Ahmed, GPT‑6.1 Sol performed well in those evaluations. He said the model must both grasp the nuance of a request and respond quickly at a cost profile the business can sustain. "I think it strikes the balance perfectly," he said.
Developer tooling and internal choices
DevRev’s engineers standardized on Codex for development work. The team explored multiple coding tools and, over time, adopted Codex across the board, discontinuing their Cursor licenses. Bashir framed this as developers choosing the best tool: "Developers are really just looking for the best tool, and you can’t fool them. ... We think that Codex is far and away the best experience."
Sales engineers also leverage GPT‑6 Astra to generate demos for DevRev’s booths at international conferences.
Operational considerations and next steps
When a customer or employee asks for help, the Operator may need to understand the issue and take an approved action, or route the task to Computer to prepare the right human to act. Ahmed noted opportunities to use the Decisions API and Ultrafast features to speed up workflows where deep reasoning is not required at every step.
Some decisions are costly to define with static rules, particularly when the dependent information changes frequently. Bashir would like builders to be able to describe more of that logic conversationally and select the appropriate level of intelligence for each task. "These things have to be able to run and complete themselves in seconds, not minutes," he said.
DevRev’s next aim is for Operator to handle longer, more demanding customer and employee requests while keeping actions traceable, permissioned, and reversible. Ahmed described routine support tasks—finding a knowledge‑base answer, upgrading a plan, issuing a refund—as only the beginning: "We want to really realize that promise of self‑service."
By combining OpenAI’s frontier models with company knowledge and control layers, DevRev is working to enable Computer and Operator to carry increasingly complex requests through to resolution on behalf of users, while maintaining accountability and traceability.
