Wiki › Engineering with agents

Modern Engineering Values

Agent-era software-engineering values: strong ownership, taste, strict guardrails, repo-local context, stack control, option value, and technical management.

Modern engineering values are the software-engineering norms Christoph Nakazawa argues still matter after coding agents make hand-written code much less central. His core claim is that programming is shifting from producing code toward directing systems that produce code, but the durable human bottlenecks remain ownership, taste, context, verification, architecture, and judgment.¹

Nakazawa's own workflow is strongly agentic: he says he rarely writes code by hand, runs Codex CLI sessions across multiple projects, starts by asking the agent to gather context and propose a plan, forces bug fixes to begin with failing tests, runs review cycles, validates functionality, and then reviews uncommitted diffs with Codiff. This is a concrete version of AI-Assisted Software Development where the human operates more like director, reviewer, and context carrier than line-by-line implementer.¹

The values he names are:

• Strong ownership: agents amplify people who understand the product and architecture, but amplify noise from people without context; he expects small teams with clear ownership boundaries to become more effective.
• Taste: because agents can generate unlimited bullshit, teams must get better at deciding what is worth doing at all.
• Strict guardrails and fast feedback loops: linting, tests, fast changed-file tooling, and automated verification determine whether an agent finishes in minutes or stalls for an hour.
• Context in the repo: because every agent session resembles a new engineer without organizational memory, specs, values, design choices, and product context should live close to the code.
• Own your stack: as implementation cost falls, teams can reconsider dependencies that outsource core product decisions and constraints.
• Option value: AI makes large changes cheaper, but a system can still be vibe-coded into a corner if architecture closes off future options.¹

This reframes Harness Engineering as an engineering-culture problem, not just a tools problem. The harness is not only prompts and CLIs; it includes repo-local context, tight constraints, fast feedback, review rituals, dependency choices, and the organizational permission to push work through without reintroducing human bottlenecks that erase agent speed.¹

The essay also complicates Cognitive Surrender. Nakazawa admits he sometimes ships code he has not studied closely when the area is less critical, tests look good, and the structure is easy to change. That is not the same as blind surrender if the surrounding feedback loops are strong, but it makes the boundary explicit: in agentic engineering, responsible trust increasingly depends on criticality, reversibility, and evidence rather than on whether the human hand-wrote every line.¹

At the organizational level, Nakazawa argues that PR and collaboration phases can destroy much of the velocity gained from agents, and that technical organizations will need to break down barriers to getting things done. He expects engineering management to become more technical, because managers who only own outcomes or direction but cannot confidently change the code will lose leverage as individual contributors become more agent-amplified.¹

Ambrosino's OpenAI Codex interview adds a parallel product-organization value: role boundaries can become more fluid without making disciplines obsolete. In his frame, AI makes it easier for PMs, designers, engineers, and other functions to cross tool boundaries, but the organization still needs people with deep command of each discipline's best practices. The agent-era value is not "everyone is everything"; it is high-agency overlap plus taste about which specialty matters when.²

Open questions:

• how teams preserve enough shared understanding when individuals can ship far more in parallel;
• when “own your stack” reduces complexity versus recreates avoidable maintenance burdens;
• how to calibrate the difference between safe evidence-based trust and Cognitive Surrender;
• whether the claimed 3–10x output gains survive broader measurement beyond one unusually agent-native engineer's workflow, especially in light of Speedup Illusion.