Suraj Gupta

Suraj Gupta is a software engineer at Warp working on agentic developer experience and AI terminal workflows. His work focuses on the systems around a coding model—its tools, instructions, context, and evaluation—and how automated software workflows can learn from previous work while remaining inspectable and affordable.

From databases to developer tools

Gupta’s early career combined software infrastructure with database research. While studying computer science at the University of Waterloo, he worked at OpsLevel on its microservice catalog, including repository integration, service ownership, GraphQL, and performance improvements. At Wish, he developed internal tooling for managing microservices and CI pipelines. A subsequent Snowflake internship involved FoundationDB, the distributed transactional database underlying Snowflake’s metadata.

His own writing explored the implementation details that determine whether an abstraction works well in practice. In a 2020 Rails tutorial, he explained how to attach computed database results to application objects without adding permanent columns or making unnecessary database round trips. Finding each user’s most-liked post provided a concrete example through which he examined query design, latency, caching, and maintenance.

At Waterloo, Gupta also researched encrypted databases under Florian Kerschbaum. He co-authored research on equi-joins over encrypted data with Masoumeh Shafieinejad, Jin Yang Liu, Koray Karabina, and Kerschbaum. Their problem was how to let a database match related records without revealing more relationships than necessary. The team developed and evaluated a scheme that limited information leakage across sequences of queries while supporting efficient hash joins.

By January 2023, Gupta was co-authoring Warp’s engineering writing with Andy Carlson. Their walkthrough of terminal internals traced an ordinary command through terminal initialization, pseudo-terminals, shell processes, keyboard input, and output rendering. His later work extended into the coding-agent systems operating through that developer environment.

Making coding agents more dependable

Gupta’s engineering work and writing examine what an agent can see, how it edits files, how it maintains a plan, and whether evaluation reflects the product developers actually use.

• Model-specific harness design: In his account of integrating Codex into Warp, Gupta explained why a capable coding model could perform poorly in an unfamiliar tool environment. Warp adapted its editing tools to the patch format Codex had learned, aligned tool names with training conventions, and removed instructions that caused execution to stall. Supporting the model required changes to both the agent harness and the client, extending the integration well beyond an API connection.

• Preserving useful context: Gupta and Daniel Peng’s work on coding-agent reliability connected evaluation results to improvements in task tracking and context management: editable task lists, summaries that retained critical instructions and task state, and file-edit responses containing the changed region rather than an entire large file. Their evaluation harness drove the actual Warp application, keeping the tested tools connected to the product.

• Selective tool discovery: Gupta contributed alongside Matthew Albright, Daniel Peng, and David Stern to Warp’s MCP search subagent. Instead of loading every external tool definition into the main agent’s context, a separate agent searched for relevant tools and resources. The main agent could then use the capabilities needed for its task without carrying definitions for unused integrations.

Improving the workflow between runs

In “How Software Factories Improve Themselves”, Gupta described leading harness development at Warp. He argued that maintaining automated workflows from issue triage through production would become an increasingly important part of software engineering, including improving the factory’s procedures, accumulated knowledge, and model choices.

His approach separates doing the work from learning how to do it better. In Warp’s software factory, an outer-loop agent examines a triage agent’s decisions and human feedback, then proposes updates to its skill through a pull request. Git records how the instructions change, and human review provides an opportunity to catch updates that could make the triage agent perform worse.

He distinguishes these procedural instructions from persistent memory: facts and outcomes retained from earlier runs. In Warp’s Oz cloud agent platform, a Sentry agent can reuse a previous investigation’s root-cause findings rather than gather the same context again. These memories work across harnesses and can be versioned, edited, deleted, and traced to their source runs, allowing engineers to inspect and correct what agents retain.

Gupta also treats model choice as a decision to test against the work a team actually performs. Warp supports automatic routing and custom rules that assign task classes to models, helping manage the cost of routine triage and simple CI fixes. He described internal evaluations that ran tasks across several models and found GLM efficient for UI work. At the time of the talk, customer-facing evaluations tailored to individual workflows were planned; his goal was to make routing decisions less dependent on intuition or generic benchmarks.