GitHub Copilot goes fully agentic

Welcome back to Building AI Agents, your biweekly guide to everything new in the AI agent field!

OpenAI’s deep research is an incredibly powerful tool which should only be used for the most profound of queries, like these:

In today’s issue…

• GitHub’s new autonomous coding agent

• Everything you need for computer use agents

• Build browser agents with an easy API

• Should we develop fully autonomous AI?

…and more

🔍 SPOTLIGHT

GitHub Copliot is acting more and more like the captain these days. On Thursday, GitHub announced a preview of agent mode, a significant upgrade to the AI coding assistant that transforms it from a passive suggester of new code to an active software engineer.

Announced in June 2021, GitHub Copilot was made generally available for public subscription a year later and gradually became the dominant LLM-powered programming copilot, widely used as a plugin to the leading IDE, VSCode. Copilot’s success spearheaded the new era of software engineering, in which programmers state what code they want and see it written almost instantly by large language models, with their role increasingly limited to validating the AI’s outputs. As Andrej Karpathy put it, “the hottest new programming language is English.”

But competition has come knocking for Copilot. Cursor, launched in 2023, offers a full, standalone IDE rather than a simple plugin. This versatility propelled it to become the fastest-growing SaaS product of all time, rising from $1 million in annual revenue to $100 million in just 12 months.

An even greater threat, however, has emerged from fully agentic coders that do not just suggest new pieces of code, but actively implement, test, and iteratively improve them. The first example to gain significant attention was Devin, which purported to take humans out of the coding loop entirely, but faced allegations of exaggerated performance. Far more successful has been Replit Agent, which won plaudits for its ability to produce entire programs in a single shot, albeit with some need for a human to check the outputs. Windsurf, a similar project launched last November, has been quickly gaining popularity as well.

Now, with agent mode, GitHub is striking back. Available now as a preview, agent mode allows users to assign a coding task to an LLM-based agent and turn over control of the workflow to it. As with other AI agents, this autonomous programmer then decomposes the job into a set of sub-tasks which it completes automatically, analyzing and correcting errors as it goes along. While the human user still monitors the agent and provides guidance or corrections, agent mode represents a radical departure from the existing paradigm of even AI-assisted coding: the human’s role has switched from engineer to manager.

GitHub also provided an initial look at an even more ambitious software engineering agent code-named Project Padawan, slated for release later this year. Unlike Copilot’s agent mode, Padawan is designed to fully automate the job of resolving bugs in code repositories. When an issue is identified, a human engineer will assign it to a Padawan agent, which can then write and test code to fix it, then task human reviewers with validating the patch, iteratively incorporating their feedback until the issue has been fully resolved.

As with each new release of AI-powered coders, GitHub’s announcement adds fuel to the debate over the future of software engineers in a world where the profession’s defining skill—the ability to write code—is outsourced to large language models. It seems likely that programming will follow a similar trajectory to many other fields and converge on a final, omni-profession: AI agent manager.

If you find Building AI Agents valuable, forward this email to a friend or colleague!

🤝 WITH ARTISAN

Hire an AI BDR to Automate Your LinkedIn Outreach

Sales reps are wasting time on manual LinkedIn outreach. Our AI BDR Ava fully automates personalized LinkedIn outreach using your team’s profiles—getting you leads on autopilot.

She operates within the Artisan platform, which consolidates every tool you need for outbound:

• 300M+ High-Quality B2B Prospects

• Automated Lead Enrichment With 10+ Data Sources Included

• Full Email Deliverability Management

• Personalization Waterfall using LinkedIn, Twitter, Web Scraping & More

Book a demo to see what Ava can do.

📰 NEWS

AI agents can now do object detection

AI researcher and educator Andrew Ng’s company Landing.AI unveiled Agentic Object Detection, which enables users to name an object and have an AI agent identify its location in an image through multi-step reasoning.

🛠️ USEFUL STUFF

Everything you need for computer use agents

A curated repository containing research, useful resources, articles, and more on computer use agents.

Browser Use’s API for easy browser agents

Browser control agent startup Browser Use has released an API that allows users to describe complex, multi-step web tasks and send them to be executed automatically by an agent.

DeepSeek’s real-time learning for your own agents

Osmosis is a YC-backed startup whose tech empowers agents to learn in real-time from their previous actions in a similar way to DeepSeek R1, outperforming OpenAI’s Operator on complex, multi-step tasks.

💡 ANALYSIS

Keeping coding agents under control

AI-powered software development is already much faster than unassisted human programming, this piece argues, but it creates concerns around trust and reliability of the created code. The balance between the two will be “human-steered agents”, which perform much of their work automatically but with close human oversight.

The law behind agent payments

Dazza Greenwood, founder of the MIT Computational Law Report, presents a fascinating deep-dive on the law governing automated transactions and its implications for agentic payments, providing agent builders with crucial advice on avoiding legal and financial risks.

AI agents are already in the building

Citing a striking (though potentially mistaken) survey result that 58% of workers are already using AI agents today, this piece discusses how to build a productive workplace around integrated agentic systems—and mitigate the trust issues that many workers are having in the tech.

🧪 RESEARCH

Fully autonomous AI agents should not be developed (?)

This position paper from Hugging Face flatly declares that humanity should not create fully autonomous agents, pointing to a plethora of risks with historical precedents, and that all agentic systems should have a measure of human control and input.

Multi-agent system design via graph neural networks

G-designer aims to solve agent builders’ endless headache as to how to architect their multi-agent systems by modeling the agents’ communication network as a graph, then using an algorithm derived from graph machine learning to design a highly-performing system.

Synthetic training data to enhance agents’ planning abilities

Creating synthetic data to fine-tune LLMs’ planning abilities is expensive and time-consuming due to the need to manually create the tasks and evaluations. The authors of this paper introduce AgentGen, which uses LLMs to automatically create synthetic tasks which LLMs can be trained on.

Thanks for reading! Until next time, keep learning and building!

If you have any specific feedback, just reply to this email—we’d love to hear from you

Follow us on X (Twitter), LinkedIn, and Instagram