DevRev bets on GPT-6.1 Sol for self-service support

DevRev aims to make its agents Computer and Operator capable of handling longer and more complex requests, while ensuring that every action is traceable, authorized, and reversible. The company relies on OpenAI models, particularly GPT-6.1 Sol and GPT-6 Astra, and claims to have found a balance between reasoning, cost, and latency for enterprise service.

Ambitious Self-Service and Controlled Actions

DevRev anticipates that Operator will be able to handle longer and more complex requests from both customers and employees, while ensuring traceability, authorization, and the ability to undo actions taken. The company designs Computer and Operator to process increasingly complex requests to completion, equipped with the knowledge and control mechanisms necessary to act on behalf of a user. Traditional support operations, such as finding an answer in the knowledge base, changing a plan, or obtaining a refund, are seen as a first step. Ahmed Bashir emphasizes that every action taken by the agents must remain reversible and that responses should reference their sources.

Business Context, Permissions, and Dedicated Evaluation

DevRev's agents rely on software that provides business context, applies permissions, and controls actions. DevRev evaluates their performance with Enterprise-Bench, a testing framework that assesses cases where data exists in multiple versions, where information is distributed across different systems, and where access depends on the user's role. OpenAI models ensure reasoning, planning, and tool selection at the core of Computer and Operator. In Agent Studio, users can describe the expected behavior of a tool in conversation and set fixed rules. Ahmed Bashir notes that GPT-6.1 Sol performs well in these evaluations.

Model Selection: Claimed Balance Between Cost, Latency, and Reasoning

DevRev uses GPT-6.1 Sol to power its agents and combines this model with GPT-6 Astra and its own system to connect internal knowledge and limit authorized actions. The company believes that OpenAI models provide the necessary reasoning and reliability while adhering to speed and cost constraints for support. Ahmed Bashir highlights a combination of suitable reasoning capabilities, trust, high alignment, and controlled latency. He believes that GPT-6.1 Sol strikes the right balance for these applications.

Execution in Seconds: Orchestrating Operator, Computer, and APIs

Depending on the request, Operator can analyze the problem, execute approved actions, or redirect to Computer to prepare the intervention of the right person. Ahmed Bashir considers that the Decisions and Ultrafast APIs offer an advantage in speeding up workflows, without the need to apply complex reasoning at every phase. He points out that some decisions are difficult and costly to formalize as strict rules, especially when information changes, and he wishes for developers to have the ability to express this logic more conversationally while selecting the type of intelligence best suited for each use. He specifies that these processes should be completed in seconds, not minutes, while considering the nuances of requests and remaining affordable for the business.

Internal Tools: Generalized Codex, Abandoned Cursor, Astra for Demos

Internally, DevRev teams have generalized the use of Codex for development. Adoption has been gradual after testing several coding tools, and the company had, at the time of the interview, abandoned its Cursor licenses in favor of a complete migration to Codex. Ahmed Bashir indicates that developers are seeking the most efficient tool and have clearly expressed their desire to continue with Codex, which the company considers to offer the best experience. DevRev's sales engineers also use GPT-6 Astra to create demonstrations to present at booths during international conferences.