Raising an Agent Season 1 - Series
The Assistant Is Dead, Long Live the Factory
A multi-episode series in which the Amp team documents building Amp. Topics include the shift from prompt to context engineering, autonomous feedback loops, sub-agent architectures for context management, the ‘Oracle’ pattern for multi-model reasoning, why codebases must adapt to agents, and evaluating models across providers. The team argues that results depend on model quality rather than hidden tricks, and treats agents as power tools rather than magic.
Episodes Browse 10 summaries 2 of 10
In episode 9 of "Raising an Agent", the hosts argue that coding agents are moving from interactive assistants into an agent-native "factory."
Key Points Covered
- Frontier models changed the baseline: Gemini 3 and Opus 4.5 reset the hosts' expectations for agentic coding. The episode argues that a smarter model can be faster and cheaper in practice when measured by completed work, not raw response time 01:08–03:06(opens in a new tab).
- Agents writing code is becoming assumed: The hosts push back on claims that agents "cannot" write real code. They argue that skepticism often lags the frontier by months. They forecast that most code will likely be agent-written, even if humans still intervene for important exceptions 03:40–05:55(opens in a new tab).
- The leash can be longer: The episode contrasts newer models with one-on-one assistants that need constant steering. It frames them as workers that can fetch context, test their output, and continue when the task is well-scoped 07:30–09:45(opens in a new tab).
- Codebases need to become agent-native: The central practical question is whether the agent knows how to use the codebase. That includes running it, verifying changes, authenticating, clicking around, inspecting output, and recovering when feedback is missing 10:18–13:03(opens in a new tab).
- Feedback loops are product surface area: Thorsten describes adding screenshot capture to a terminal emulator and a CLI subcommand for Amp TUI data. These let an agent verify visual or UI work without slow human screenshots or brittle tmux control 15:10–17:24(opens in a new tab).
- The development factory changes team process: If agents are cheap to launch, writing tickets, waiting for sprint planning, and reviewing expensive human-authored pull requests start to look like bottlenecks. The hosts expect more immediate investigation, more parallel variants, and more trust in senior engineers reviewing agent-produced work 25:29–31:43(opens in a new tab).
- Amp is optimizing for frontier users: The team says model selectors and cheap-model optimization would slow Amp down or select for the wrong users. It says subscription-style token bundles would do the same when the product needs to keep changing with the frontier 20:36–24:56(opens in a new tab), 32:18–37:40(opens in a new tab).
- Deep mode is teased as the next factory primitive: The episode hints at a mode for longer-running work. The agent goes off, researches, and returns with a result instead of asking the human for each next step 38:43–40:38(opens in a new tab).
Full video: https://www.youtube.com/watch?v=2wjnV6F2arc(opens in a new tab)
Episode page: https://ampcode.com/podcast/episode-9(opens in a new tab)