Raising an Agent Season 1 - Series
Killing the Sidebar
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 1 of 10
When agents can work for longer without input, a single editor sidebar becomes a poor place to manage several tasks. In this February 2026 episode of "Raising an Agent", Quinn and Thorsten explain Amp's move away from that interface and toward longer-running work with later human review.
Key Points Covered
- Deep mode is a workflow, not just a model selector: At the time of recording, Amp's new deep mode used GPT-5.2 Codex Medium for longer, well-scoped tasks that benefited from thorough research and delayed review. The hosts contrast it with smart mode for interactive work and rush mode for small fast changes 00:55–05:21(opens in a new tab).
- Modes should change user expectations: Quinn and Thorsten argue that putting a model in a drop-down is not enough. A product should help users feel when they are writing a short text-message-style prompt versus a longer letter to an agent that will go away and work 05:52–07:31(opens in a new tab).
- The factory pattern becomes more natural as models run longer: If an agent can work for 45-60 minutes without needing input, watching one sidebar thread stops making sense. The hosts expect more parallel agents, long-running tasks, and check-ins after completion 08:03–10:17(opens in a new tab).
- Skills are accumulating operational knowledge: Amp uses skills to encode release workflows, tmux testing details, G-Cloud log analysis, and BigQuery usage. These and other team-specific procedures help agents navigate common tasks without repeated prompting 10:17–16:18(opens in a new tab).
- Agent-optimized dev tooling can change human workflows: The team discusses fast cached test commands. It also discusses accepting some regression in VS Code ergonomics when that tradeoff improves the feedback loops agents use most often 16:18–21:19(opens in a new tab).
- Amp is killing the editor extension: The hosts say the sidebar limits parallelism and keeps the human as the feedback loop. They say supporting it as a first-class path would select for lagging users. They urge users to move to the CLI while the VS Code extension self-destructs 21:52–27:57(opens in a new tab).
- Software product assumptions are melting: The hosts cite AI-generated workout UIs, voice-guided workflows, and assistant-managed shopping lists. They argue that software may become less tied to durable app interfaces and more tied to generated or conversational workflows 28:31–32:52(opens in a new tab).
- Frontier companies need to keep re-earning their product: Amp lacks a traditional sales/marketing buildout and is willing to remove features and make risky changes. It frames these choices as a strategic response to models that can overturn product assumptions every few months 33:25–43:49(opens in a new tab), 49:52–55:05(opens in a new tab).
- Manual context management may be fading: The episode revisits an old assumption. As models such as GPT-5.2 improve at researching after compaction, automatic handoff and hidden context-window management become more viable. However, clear task boundaries still matter 43:49–49:19(opens in a new tab).
Full video: https://www.youtube.com/watch?v=4rx36wc9ugw(opens in a new tab)
Episode page: https://ampcode.com/podcast/episode-10(opens in a new tab)