Freelance Agent Evaluation Engineer

Overview

• We're building a dataset to evaluate AI coding agents how well a model handles real-world developer tasks.
• You'll create challenging tasks and evaluation criteria within realistic simulated environments:
• Build virtual companies following a high-level plan - codebase, infrastructure, and context (conversations, documentation, tickets) that form a realistic environment with development history
• Assemble and calibrate tasks from intermediate states of the virtual company: craft the prompt, define evaluation criteria, and ensure the task is solvable and the evaluation is fair
• Design tasks set in isolated environments - emulations of a developer's workstation: a Linux machine with development tools (terminal, CLI), MCP servers (repository, task tracker, messenger, documentation, etc.), and a real web application codebase
• Write tests that accept all correct solutions and reject incorrect ones - neither too strict (breaking on valid approaches) nor too lenient (passing bad ones)
• Iterate with an AI agent on tests - verifying they catch real problems, don't miss bad solutions, and don't break on good ones
• Review code written by agents, analyze why an agent failed or succeeded, and design edge cases and adversarial scenarios
• Iterate based on feedback from expert QA reviewers who score your work on quality criteria
• A significant part of the work is done together with AI - it's very hard to create tasks that challenge frontier models without using frontier models.
• Why this is hard
• Frontier models are already good at coding. Creating a task that genuinely challenges the best models is non-trivial. You need to deeply understand where models fail and what scenarios reveal the difference between a good and a bad solution.
• Tasks have many valid solutions. Writing tests that accept all correct solutions and reject incorrect ones is harder than it sounds.