Show Notes
In this episode, Beyang and Thorsten discuss strategies for effective agentic coding, including the 101 of how it's different from coding with chat LLMs, the key constraint of the context window, how and where subagents can help, and the new oracle subagent which combines multiple LLMs. 00:53 Intros 03:35 How coding with agents is very different from coding with prior AI tools that use chat LLMs 10:46 Example of an agentic coding run to fix a simple issue 14:28 Example of debugging an issue with an MCP server 22:05 Example of unifying two build scripts that share logic 25:24 How context window size has emerged as a key constraint on agentic automation 31:16 Why it's best to focus on one thing at a time per agentic thread 33:24 Subagents and how they help extend the effective context window 34:04 The Amp codebase search subagent 38:48 General-purpose subagents 44:20 When to use subagents 47:04 The oracle subagent and o3 51:47 Multi-model agents and using the best model for each job
More from Raising An Agent
Quinn and Thorsten sit down together in San Francisco to answer the question they get asked most: how does the Amp team work? They walk through a week at Amp, from Thorsten's Sunday post and the single Monday meeting to writing down goals in silence. They explain why Amp has no quarterly goals, no departments and no territory, and how feature flags and the work in progress, orb good and orb bad channels keep a team of 21 in sync. They also cover reviewing changes with agents, pushing to main without mandatory code review, and why they think more companies will soon work like this.
Quinn and Thorsten record a late-night Hot Take Thursday. They start with a company that capped each engineer at €50 of tokens a week, then talk about why cheap models end up costing more and why some people will be priced out of frontier models. They explain how agents now ship, monitor and fix production, not just write code, and why code review and other old processes get in the way. They also cover not waiting on your agents, simple evals in Orbs, why Amp has runners, and the hot takes Thorsten left out of his post.
Quinn and Thorsten discuss what they can hand off to agents now, from coordinating work across projects to following up with users after a fix ships. They explain how they test agent behavior and use production data to improve performance, then talk about Amp's recent outages and the work to improve reliability. With more of their work happening in Orbs, they ask what the personal computer is still for, and why changing how you work matters as much as changing models.
Quinn and Thorsten predict that intelligence will get cheap and tokens will be everywhere, then ask what that does to software. They debate build versus buy in the age of agents, why Orbs are worth paying for while Thorsten one-shot a bug tracker, and why developers should do X instead of building software to do X. They also explain the two Amps, why building Amp feels like owning a house instead of renting, and how they plan to bring the Amp way to more teams.
This episode was recorded in person in Munich, where the whole Amp team had gathered for a week of building. Quinn and Thorsten discuss why isolated Orbs make it easier to run agents in parallel, how the team asks agents for exhaustive verification, and how production itself can become a feedback loop. They also introduce jellyware: high-quality software that is designed to change quickly. When agents make software more malleable, teams can fix bugs faster and raise the quality bar instead of lowering it.