GitHub Released Deep Research and Didn’t Even Know It (It Even Beat Perplexity)

GitHub Copilot Coding Agent (GCCA) as Deep Researcher?

TLDR: GCCA is a powerful autonomous agent capable of addressing GitHub issues and resolving them by reviewing, writing, and testing just like a human developer would. It can also use its powerful agentic framework for many other tasks. In a sample of 1 (very scientific) comparison, I ran the same report subject through GCCA and Perplexity, then had my AI review agent subjectively determine a winner—GitHub won 3-2.

Deep Research

I’m keeping this short and sweet because I have a few other interesting findings from this week to share. Check out the public repo links below and take GitHub Deep Research for a spin.

The Deep Research capability first rolled out in December 2024 when Google debuted it as part of their Gemini AI offering. By the end of February 2025, we had OpenAI, Perplexity, and a host of other AI companies and open-source projects debuting their offerings. While GitHub doesn’t officially have a deep research offering, I decided to test it on problems I would normally use a Deep Research tool for.

Using GitHub Copilot Coding Agent for deep research is straightforward. Create a GitHub issue and outline in the body what you’re looking for as an end result. That might be a single markdown file (like I used in the technical research request test) or a more elaborate file structure that breaks out various parts into different files and folders, then provides a cross-linked report. The agent is your oyster—ask and it will obey. To get GCCA involved, simply assign it to the issue. You do need a paid GitHub Copilot account to use GCCA.

The Problems:

1. General technical research (the current debate over foundational LLMs vs. a Mixture of Agents architecture)
2. Planning a trip (ski trip to Stowe)

Problem #1

I assigned both GCCA and Perplexity the same prompt:

“What is the latest research on the use of large SOTA foundation models vs. a Mixture of Agents approach using smaller open-source models?”

I then took the single markdown-formatted report from both and created a new issue in GitHub, having GCCA evaluate and score the results. As a GitHub employee, I’m too biased. GCCA may be biased, but I did tell it not to be in the prompt (for whatever that’s worth). Here’s how it turned out:

GitHub ekes out a win. Check out both reports and the results in the GitHub Repo.

Final Scoring Summary

Problem #2

This one was more straightforward and didn’t include a comparison. I created a GitHub issue and had it plan a family vacation to Stowe. It failed the first time because I forgot to adjust or disable the firewall that sandboxes GCCA from reaching the internet. I opened it up so it could perform a broader set of searches than it would normally be able to do. The second run worked really well—the plan was much more thorough than I could have expected. It did fail to pull screenshots (even with the Playwright MCP server hooked up) of the Airbnb options it recommended. I’ll leave it up to you to judge how it did—check out the GitHub Repo.

This is not a scientific study, just me screwing around as usual. Good luck Researchers!!