AI as Research Copilot

Created on April 14, 2026

2026   ·   AI   research-workflow   Claude-Code   ·   research

By 2026, AI is no longer just a coding tool for me. It’s part of the research loop: discussing ideas, iterating experiments, writing reports, and reflecting on feedback.

The tools I tried

In 2024, I tested Trae, Cursor, and Copilot. Cursor was the first that felt genuinely useful during ForgeHLS development. Codex became my go-to for straightforward coding—cheap, reliable, no quota anxiety.

Claude Code: the real shift

Claude Code changed everything. It can sustain long chains of work: research ideas → scope → implement → analyze → deliverables. I use it in auto-research-single for literature research and benchmark-grinder for experiment iteration.

The weakness is cost. But it’s the first tool that feels like a research copilot, not just a code generator.

My workflow now

I use AI across the full research pipeline:

  • Idea scouting: Pressure-test ideas, find baselines, expose assumptions
  • Experiment iteration: Modify experiments, compare outputs, debug—the biggest gain
  • Deliverables: Turn work into reports and papers
  • Meeting reflection: Unpack feedback and plan next steps

I formalized this into reusable agent skills like autonomous-benchmark-optimizer and benchmark-grinder.

What I pay for

  • Codex: ~$20/month for straightforward coding
  • Cursor: ~$20/month
  • Claude Code: main workhorse via third-party relay (official is too expensive)
  • OpenRouter: for LLM experiments

The real change

AI didn’t just make me code faster. It made my research process externally runnable. Parts that lived only in my head—idea formation, experiment iteration, report writing—now happen in an interactive loop.

The uncomfortable truth: more research work looks automatable than I expected. Literature mapping, baseline design, experiment iteration, technical writing—AI already handles much of this. From where I stand, research feels less like a safe exception to automation and more like a domain being quietly reorganized by it.

AGI may arrive not as a dramatic event, but as a steady collapse of tasks that once defined skilled intellectual work.