Raising An Agent - Episode 10

Core Argument

The one-on-one sidebar assistant is reaching its ceiling. Models like GPT-5.2 Codex Medium work best when sent off autonomously for 45+ minutes, not babied through ping-pong interactions. AMP is killing its VS Code extension to force this shift: from watching a single agent in an editor to orchestrating a factory of parallel agents from the CLI.

Key Takeaways

• Agent "modes" are not model selectors. Deep mode (GPT-5.2 Codex Medium) is lazy at quick assistant tasks but extraordinary at scoped, long-running research. Smart mode (Opus 4.5) excels at fast interactive tasks. Rush mode (Haiku) handles trivial work. Each mode reshapes the UX, not just the backend.
• Skills are the winning abstraction. Unlike MCPs, custom commands, or sub-agents, skills—small instruction documents encoding domain-specific knowledge—have durably stuck. AMP uses them for GCloud log analysis, BigQuery queries, and release procedures. The compounding effect accelerates: every skill added makes the next task cheaper to delegate.
• Optimizing for agents degrades human DX—and that's the point. AMP built Zveltch Check, a Zig-based Svelte type checker that made agents faster but broke VS Code integration. They chose the agent experience, creating a self-reinforcing flywheel: degraded human paths push more work through agents, which justifies further optimization.
• AMP is killing its VS Code extension. The sidebar limits parallelism, acts as a crutch, and attracts users who want an assistant rather than a factory. A 60-day sunset forces both the team and users onto CLI-based multi-agent workflows.
• Manual context management is dying. Auto-compaction—automatic summarization of context windows—makes long-running agents viable. Agents re-research lost context after compaction rather than requiring humans to manage thread state.

The Factory Model

flowchart LR
    subgraph Sidebar["Sidebar Era (dying)"]
        H1[Human] <-->|"watch & wait"| A1[Single Agent]
    end
    subgraph Factory["Factory Era"]
        H2[Human] -->|"kick off tasks"| CLI[CLI / Headless]
        CLI --> D1[Deep Agent 1]
        CLI --> D2[Deep Agent 2]
        CLI --> D3[Deep Agent N]
        D1 & D2 & D3 -->|"45 min later"| R[Verified Results]
    end
    Sidebar -.->|"evolution"| Factory

AMP's Three Agent Modes

Predictions Made

• The sidebar will be obsolete within months — AMP sets a 60-day self-destruct timer on their VS Code extension
• Agent parallelism is the primary productivity lever — not smarter single-agent interactions
• Auto-compaction replaces manual thread management — models will re-research context naturally
• Cursor and editor-bound tools will be forced out of the editor — the fastest-growing startup is already hearing "VS Code is holding you back"
• Traditional SaaS playbooks break down when the market shifts every 2-6 weeks

Notable Quotes

"I think the time of you one-on-one with an agent in a sidebar going back and forth — I think that's coming to an end." — Torsten

"For the 1% of developers that want to be most ahead, that want to be coding like how everyone else will be in the future... they only need to do the last 20% of their work in the editor and we think we can get that to 10% or 1%." — Quinn

"We had this difficult decision — which is more important: to preserve the human dev experience or to preserve the agent dev experience?" — Torsten

"We want to be the most radically on the frontier one. We want to rip things out way faster than anyone else. We want to carry less baggage than anyone else." — Quinn

Resources Mentioned

• Zveltch Check — Open-source Zig-based Svelte type checker built for agent speed
• Pi — Minimal agent framework by Mario Zechner, referenced as exemplar of radical independence
• Peter Steinberger — Built OpenClaw/ClaudeBot on Pi, vocal advocate for GPT-5.2 Codex
• Sourcegraph — Company AMP spun off from (all 19 AMP employees are ex-Sourcegraph)

Connections

• raising-an-agent-episode-9 — Previous episode in the same series; introduces the "factory" mental model that Episode 10 doubles down on by killing the sidebar
• the-importance-of-agent-harness-in-2026 — AMP's modes and skills are a concrete implementation of the agent harness architecture this article describes
• pi-coding-agent-minimal-agent-harness — Mario Zechner's Pi agent is directly mentioned as an exemplar of the minimal, autonomous approach AMP champions
• anthropic-just-dropped-agent-swarms — Claude Code's agent teams implement the same parallel multi-agent pattern AMP advocates with their factory model