← back
The emerging skillset of wielding coding agents — Beyang Liu, Sourcegraph / Amp
Original: The emerging skillset of wielding coding agents — Beyang Liu, Sourcegraph / Amp
Takeaway
Coding agents need a new operating mode (direct edits, oversight not approval, usage-based pricing, model-coupled scaffolding) — old chatbot intuitions actively hurt.
Summary
- Coding agents (post Claude Sonnet/GPT-class tool-use) are a paradigm shift from chatbot+RAG; best practices from 6 months ago are now wrong.
- Sourcegraph's Amp principles: agent edits files directly, no per-turn approvals; users should sit above the loop, not micromanage.
- Skeptical of forked VS Code shells and fixed-price plans — usage-based pricing aligns incentives with smarter models, gym-membership pricing pushes toward dumber/cheaper models.
- Models become tightly coupled to agent scaffolding; swappable-model assumption breaks down because tool-use quality varies dramatically.
- Jeff Huntley's Canva interviews: most engineers were 'holding it wrong' by using agents like autocomplete.
coding-agentssourcegraphdeveloper-experience
Original description
It's raining coding agents! But while many are saying they're feeling the AGI, others say they're not that useful for serious programming. How much is hype and how much is a skill issue? We'll share empirical observations that help explain the divergence of developer opinion. And we'll cover emergent strategies uncovered by users of Amp, a new coding agent in research preview, that can help you employ agents to complete more complex tasks in production codebases. About Beyang Liu Beyang is the co-founder and CTO of Sourcegraph, the company behind Sourcegraph Code Search and Amp. Beyang started his career working on software for some of the largest banks as an engineer at Palantir, where he brought a background in machine learning and data analysis at Stanford. Recorded at the AI Engineer World's Fair in San Francisco. Stay up to date on our upcoming events and content by joining our newsletter here: https://www.ai.engineer/newsletter Timestamps Introduction & The State of AI in Coding [00:14] Current Discourse: The talk begins by acknowledging the polarized debate on AI's role in coding. While some elite programmers are skeptical, many developers find significant value in AI tools, suggesting a disconnect between top-tier and mainstream experience. Liu frames the discussion by referencing opinions from figures like Jonathan Blow and Eric S. Raymond, highlighting the varied perspectives in the field. [03:01] Paradigm Shift: The most significant mistake developers make is using new agents with old mental models. Liu emphasizes that we are in a "step function transition" in model capabilities, meaning that strategies from even six months ago are already outdated for leveraging the full power of today's agents. The Three Eras of AI Coding Tools [05:06] GPT-3 Era (2022): This era was defined by text completion models. The primary application was "copilot" or "autocomplete," where the AI would suggest the next few lines of code based on the preceding context. [05:24] ChatGPT Era (2023): The introduction of instruct-tuned models like GPT-3.5 led to the rise of chatbots. In the coding world, this manifested as "ragbots," which combined a chat interface with a retrieval engine to answer questions about a codebase. [06:11] Agent Era (Present): The current era is defined by models capable of tool use and autonomous operation. This requires a new application architecture where the agent can directly edit files, run commands, and interact with external services to accomplish a goal. Controversial Design Philosophy for Agents [07:27] Autonomous Edits [09:55] Unix Philosophy [10:24] New Applications Live Demo: Sourcegraph's Amp Agent [13:15] The Task [14:30] Tool Use [15:53] Sub-Agents [17:56] Planning & Execution [19:46] Nuanced Problem Solving Best Practices from Power Users [23:21] Detailed Prompts [24:21] Feedback Loops [28:03] Code Understanding [28:36] Code Reviews Anti-Patterns and Future Outlook [30:35] Micromanagement [30:46] Under-prompting [31:52] Parallel Agents [33:18] High-Ceiling Skill