Talha Sheikh is an AI software engineer and creator of Vector Harness, a developer tool that checks whether coding agents actually satisfy the requirements they claim to have completed. His work tackles coding-agent reliability: software-generating models can produce plans, tests, and confident completion reports while still leaving developers responsible for discovering what does not work.
Sheikh studied at the University of Westminster from 2017 to 2020 and was affiliated with Checkout.com at AI Engineer Europe 2026. His other work has included AI interview-practice tools, hackathons, and experiments with assistants designed to reduce mental fatigue. He developed Vector during his London commute, documenting his experiments with remote coding agents and the practical obstacles to supervising them effectively.
- Deterministic verification: Vector uses Claude Code hooks to run configurable, pass-or-fail quality gates against generated work. Failed checks send feedback to the agent for another attempt; persistent failures escalate instead of silently passing.
- Capability is not reliability: Larger models, detailed specifications, additional context, and elaborate agent orchestration can improve what a system is capable of producing without proving that its output fulfills a developer’s actual requirements.
- Developer-defined enforcement: Sheikh shifted Vector away from prescribed checks toward a configurable contract that lets each team determine what successful execution means across programming languages and workflows.
- Verification throughout the workflow: Checks can operate when an agent finishes, before commits, between stages of multi-agent workflows, during asynchronous execution, or alongside model-based evaluation. Sheikh also argues that stronger guardrails could make smaller models viable for some tasks, although the practical savings depend on the workload.
At AI Engineer Europe 2026, Sheikh framed harness engineering as a shift in where software developers create value: defining requirements, designing feedback loops, and independently determining when an autonomous coding task is genuinely complete.
