Raising An Agent - Episode 9

Core Argument

"The assistant is dead, long live the factory." Agents writing code is now table stakes. The differentiator is whether your codebase is built for agents—with feedback loops, testability, and verification baked in.

Key Takeaways

Model capability outpaces expectations. Opus 4.5 and Gemini 3 changed the game. Skeptics (including Karpathy) reversed their views after trying frontier models in modern harnesses.
Smart models are faster AND cheaper. Measuring "task completion" rather than token cost reveals that frontier models outperform fast-but-dumb alternatives.
Longer leash, not tighter control. Instead of back-and-forth assistant interactions, give agents full autonomy to investigate, implement, test, and iterate.
Agent-native codebases are competitive advantage. The 2026 equivalent of "can someone ship on day one?" is "can an agent ship in 10 minutes?"
Build feedback loops INTO applications. Example: Adding --capture CLI flags to applications so agents can screenshot their own output and verify visually.
Dev tooling assumptions are broken. Linear tickets, pull request reviews, emoji reactions—all built for humans who invested effort. Agents change that equation.

The Factory Mental Model

flowchart LR
    subgraph Old["Assistant Era"]
        H1[Human] <-->|"back & forth"| A1[Agent]
    end
    subgraph New["Factory Era"]
        H2[Human] -->|"queue tasks"| F[Agent Factory]
        F --> A2[Agent 1]
        F --> A3[Agent 2]
        F --> A4[Agent N]
        A2 & A3 & A4 -->|"verified output"| CB[Codebase]
    end
    Old -.->|"evolution"| New

Making Codebases Agent-Native

The Joel Spolski checklist for 2026: Can an agent ship something in the first 10 minutes?

Requirements for agent-ready codebases:

Feedback loops - Agents must verify their own changes
CLI accessibility - Expose data through commands, not just UIs
Automated testing - CI that agents can run and interpret
Documentation - AGENTS.md explaining how to use the codebase
Sandboxed execution - Isolated environments for parallel agent work

"Think of your codebase as an application. Does the agent know how to use it?"

Notable Quotes

"Writing code is over... Opus 4.5 can do a file."

"If you're using a coding agent and it's not pushing you to make changes to everything in how you build software, then it's probably not pushing you enough."

"Instead of creating a Linear ticket, we just send off an agent."

On Model Selection

AMP intentionally has no model selector. The reasoning:

Forces everyone onto the frontier
Prevents optimizing for outdated capabilities
Reduces combinatorial complexity
Enables learning from unified usage patterns

"If you're not working on an agent, do you really want to spend your time exploring the strengths and weaknesses of seven different models?"

Connections

12-factor-agents - Shares the philosophy that agents need focused context and clear feedback loops
the-importance-of-agent-harness-in-2026 - Directly discusses the harness architecture Amp is building
building-effective-agents - Anthropic's complementary guide on composable agent patterns