Amp Code Founder on future of Coding Agents

Quinn Slack describes Amp's split from Sourcegraph, changing coding-agent and SaaS economics, and Amp's focus on frontier users.

Key Points Covered

• Sourcegraph's path to Amp began with code understanding: Slack traces Sourcegraph from large-codebase search toward a long-term goal of automating more software work. This depended on code being indexed, understood, and manipulated at scale [00:04:03]-[00:06:02].
• AI shortens the loop between builder and customer: He argues coding agents let people with taste build closer to the user and market feedback. He warns that easy code generation can also produce slop [00:06:50]-[00:07:54].
• Amp reacted against enterprise-safe assistant habits: Sourcegraph's Cody work taught the team that fast-moving AI did not fit old enterprise procurement, roadmap, and policy expectations. That approach lost to frontier-first products like Cursor [00:09:50]-[00:13:37].
• Amp is not meant for every developer: Slack says Amp is for users who want the best frontier agent, even if it costs more. He says many companies should use it for their most agent-forward developers rather than standardizing every seat on it [00:13:37]-[00:14:40].
• "Coding agents are dead" means the category is escaping its box: He argues that more general agent work and model-lab pricing are disrupting the old sidebar assistant. He says they are also disrupting the narrow coding-agent category and flat-subscription business model [00:15:29]-[00:18:05].
• Model-lab subscriptions changed expectations: Cheap Claude and OpenAI subscriptions helped developers experience agentic coding. Slack says third-party agent companies cannot copy those economics when they must pay API token costs [00:20:00]-[00:24:45].
• Small-team speed is Amp's strategic bet: Amp spun out as a small team so it can move faster than both large labs and larger enterprise software organizations [00:21:55]-[00:22:48].
• Human review depends on risk: Slack says humans should scrutinize agent output for core API and data work. For low-risk internal tools, he says functional code can be acceptable even if not idiomatic [00:25:47]-[00:30:37].
• Cloud agents are the harder next product problem: He says local CLIs are easy compared with running many isolated agents in the cloud. Those agents need dev servers, screenshots, videos, and reviewable feedback loops [00:32:17]-[00:33:15].
• "Software is dead" targets the SaaS model: Slack argues that durable SaaS assumptions weaken when customers can generate the specific 10-20% of product they need. He says software companies also face token costs, talent competition, and shorter product lifetimes [00:33:15]-[00:38:58].
• Amp rejects moat-first thinking: He says no AI product has a moat proportional to what it needs to protect. He says Amp's practical difference is its willingness to move with frontier models, be expensive if necessary, and kill features that hold users back [00:45:57]-[00:53:09].

Full video: https://www.youtube.com/watch?v=_L8xxUXOTk0(opens in a new tab)