The Truth About Coding Agents: Why 90% of Your Time Is Now Code Review

Beyang Liu argues that as agents write more code, understanding their changes becomes a developer's main constraint. In this a16z conversation, the Sourcegraph CTO and co-founder discusses review interfaces, evaluating agent behavior, and the open-model ecosystem. The title's “90%” reflects practitioner reports about time spent reviewing, not a measured industry-wide allocation 29:28–33:30(opens in a new tab).

Sourcegraph’s New Agent: Amp

Sourcegraph built its coding agent, Amp, from first principles once tool-using models became reliable enough for programming work. The company kept Amp separate from existing code-search or RAG products. This allowed it to design around models that edit files, run tools, and recover from feedback 05:39–07:30(opens in a new tab).

• Enterprise and hobby use cases: Amp inherits Sourcegraph's experience with large enterprise codebases, but Liu also emphasizes hobbyists and non-programmers. One example is his dad using agents to build iPad math games, showing that the tool is not only for professional software teams 07:30–08:18(opens in a new tab).
• Smart and fast agents: At the time of the interview, Liu describes a smart paid agent for harder work and a fast, ad-supported path for cheap targeted edits. The distinction reflects a tradeoff between response time, capability, and cost 08:18–11:25(opens in a new tab), 20:09–21:56(opens in a new tab).
• The agent is the product abstraction: Liu argues that users should think about the agent, not only the raw model. Behavior comes from the model plus the system prompt, tools, tool descriptions, and feedback-loop instructions 13:23–15:15(opens in a new tab).
• Specialized workflows matter: Sourcegraph uses closed and open models. It sees open-weight models as particularly suitable for post-training specialized workflows such as context retrieval, library fetching, and targeted edits 21:56–27:35(opens in a new tab).

The Shift From Coding to Orchestrating

• Review needs more than a file-by-file diff: Liu wants tools that reveal the important decisions in a large change. The speakers suggest grouping edits by task and using explanations or diagrams, rather than making reviewers repeatedly expand isolated diff fragments 28:25–33:30(opens in a new tab).
• Orchestration still requires comprehension: Liu expects developers to direct multiple agents while remaining able to understand their outputs. People still supply the intended outcome and choose between product and system tradeoffs 29:28–31:35(opens in a new tab).
• Higher output can make work less enjoyable: Liu reports developers saying they have never been more productive but now spend most of their time on code review, which they find less satisfying than building 30:32–32:32(opens in a new tab).

Open Models, Tool Use, and Geopolitics

The speakers worry about US application builders becoming dependent on Chinese open-weight models. They discuss whether copyright lawsuits, developer-liability concerns, and uncertain state-by-state rules discourage US model releases. They present this as a possible explanation; Liu says he lacks inside knowledge of the research organizations 33:30–44:11(opens in a new tab).

Philosophy on Agents and Reliability

• Different execution paths can still reach a useful result: The hosts raise the difficulty of delegating logic to a model whose behavior varies. Liu uses code search as his example: an agent may choose different searches each time yet still find the needed context. His confidence is a report about that workflow, not a guarantee for arbitrary agent tasks 15:15–17:59(opens in a new tab).
• Use evaluations to catch regressions, not define the whole product: An eval tests an important workflow so a change that breaks it becomes visible. Liu warns that such tests can lag what users value. He cites autocomplete acceptance rate: optimizing for accepted suggestions did not establish that the code was ultimately useful, correct, or merged 17:10–20:09(opens in a new tab).

Policy and Product Implications

• Better orchestration interfaces: Liu expects the future IDE or CLI to become less of a text-editing surface and more of a control room for launching, monitoring, and understanding multiple agents 28:25–33:30(opens in a new tab).
• Regulate applications without locking the model layer: His policy recommendation is clear national regulation focused on applications, competition at the model layer, and avoiding rules that entrench incumbents or slow open-model progress 44:11–45:14(opens in a new tab).

Full video: https://www.youtube.com/watch?v=Jxz4GJSG8ZA(opens in a new tab)