AI 能编写更多代码，但系统工程仍需工程师负责

When people say AI has transformed software development, I want to know which part they mean. Producing an implementation? Exploring a design? Deciding whether a system solves the right problem?

Those activities are connected, but automating one does not settle the others. My argument is that agents can take on more implementation work, and assist with analysis, design, and testing, while engineers remain responsible for deciding what to build and what evidence is sufficient to accept it.

That changes the work. It does not make the responsibilities disappear.

I am wary of describing this as replacing the person who churns out code. It reduces developers to typing, understates what agents can assist with, and turns task automation into a claim about jobs. Claims about workforce replacement require evidence beyond an impressive demonstration.

The practical question interests me more: how do we use these tools while keeping our decisions visible and our feedback useful?

Give decisions somewhere to live

A conversation is convenient for exploring options. It is a fragile place to keep the only record of an architectural decision.

Model context is bounded. A long session may be summarized, compacted, or resumed with selected information. The summary can preserve the conclusion and lose the condition that made it reasonable. “Use a local cache” is different from “use a local cache because this screen must remain usable without connectivity.”

Anthropic's account of long-running agents describes continuity problems across context windows and an approach using incremental work and persistent progress artifacts. It also reports that compaction can leave unclear instructions. This is a vendor engineering report about a particular setup, rather than a universal comparison of development workflows.

My preference is to give durable decisions a small, reviewable home. A PROJECT_GUIDE.md might contain:

• The product's purpose and current scope, including explicit exclusions.
• The important system boundaries and the reasons behind architectural choices.
• Quality expectations and the commands used to check them.
• Links to deeper decisions when those decisions need more space.

The filename has no special power. An agent will not necessarily load it automatically. I would connect it through the tool's supported project instructions or explicitly include it in the task, then check the agent's interpretation.

For each increment, I would supply the relevant files and ask for a brief account of the intended change, constraints, and acceptance cases. A paraphrase is an early opportunity to catch misunderstanding; the implementation still has to demonstrate compliance.

Simon Willison's June 2025 discussion of context engineering helps frame the problem of choosing information for the next model step. My narrower recommendation is to make that selection deliberate and inspect its consequences.

Markdown alone guarantees neither retrieval nor obedience. Sending every document on every task can increase context and cost. Selective retrieval may reduce input, but maintaining documents, retrieving them, and correcting mistakes also consume resources. Any claim of savings needs a measurement of the complete task.

Specify the next behavior

Consider a hypothetical mobile search screen. This is an illustration, not a report from a completed project.

The story could be: “As a user, I want to search saved items so that I can find an item without scrolling through the entire collection.” For the first increment, the scope is local search over existing items. Remote search and changes to persistence are excluded.

Before implementation, I would make these acceptance cases explicit:

• A matching query displays the expected items.
• A query with no matches displays an empty result state.
• Clearing the query restores the complete collection.
• If filtering is asynchronous, an older result cannot overwrite a newer query's results.

These cases expose a technical question: does filtering need asynchronous work at all? The agent can help assess that decision. We should resolve it against the expected data size and responsiveness requirements, rather than introducing concurrency because a template expects it.

For this example, I might choose unidirectional data flow, explicit UI states, and a separation between filtering logic and rendering. That makes state transitions easier to inspect. Another system may justify a different structure; this is an engineering choice with a purpose.

The project guide should preserve that purpose. Feature-level notes should explain the behavior and any new decision. They do not need to reproduce the entire architecture.

After the increment, I would inspect the diff, run the relevant checks, and exercise the screen. If the agent rewrote persistence despite the scope restriction, passing tests would not make that expansion acceptable. If the empty state looks correct but the search matches the wrong field, a screenshot would not establish correctness either.

Make generation answer to verification

Agent output is probabilistic. Acceptance should depend on evidence that we can inspect and, where practical, reproduce.

A local verification command can run formatting checks, static analysis, compilation, and focused tests. These checks answer different questions. A formatter checks conventions; a compiler checks language and type rules; tests evaluate selected behavior under selected conditions. None establishes that we chose the right product behavior.

Controlled verification also needs controlled inputs. Tool versions, fixtures, and execution conditions matter. Tests involving timing, devices, networks, or external services may be flaky. Calling a test suite deterministic does not make it so.

Even a perfectly repeatable check can verify the wrong thing. A search test might assert that some result appears without checking which result it is. A generated test can faithfully encode the same misunderstanding as the generated implementation. I would review the assertion against the acceptance case and use deliberately chosen examples, including a failure case.

Running the app adds another kind of evidence: whether the behavior makes sense through the actual interface. When performance matters, define the measurement before declaring success—device, dataset, scenario, and threshold. A quick impression of responsiveness is useful feedback, but it is not a latency result.

Hooks can make local checks convenient and catch mistakes earlier. Their coverage depends on what triggers them and whether they can be bypassed. Where a team requires enforcement, server-side checks and repository protections need their own configuration. A local hook is not equivalent to an enforced remote policy.

The engineer still decides whether the evidence supports acceptance, revision, or rejection. An agent can help gather that evidence; it cannot make an incomplete test suite complete by announcing success.

Documentation has to earn its maintenance

I become skeptical when a small change requires a long chain of generated documents before anyone can see working behavior. Each document becomes something to review, reconcile, and keep current. Producing text cheaply does not make understanding it cheap.

That is a criticism of proportionality and feedback cadence. Spec-driven development has multiple meanings. Birgitta Böckeler's 2025 examination of Kiro, spec-kit, and Tessl distinguishes approaches that write specifications first, retain them during maintenance, or treat them as the primary source. Those differences matter when discussing costs and control.

François Zaninotto's critique of document-heavy workflows raises concerns about review overhead and instruction following through reported examples. Those cases are useful prompts for investigation. They do not establish that specifications generally cause worse outcomes.

An incremental specification can support rapid feedback. Agile does not mean avoiding documentation: its principles emphasize frequent working software, technical excellence, and adaptation. A short acceptance case can serve those aims; so can a detailed design when the risk warrants it.

Stronger specifications can be valuable for safety constraints, regulated work, public interfaces, migration compatibility, and coordination across teams. In those settings, a concise guide is an entry point to the necessary detail. It cannot replace that detail.

My test for a document is practical: what decision, implementation, or verification does it support, and who will maintain it? I would keep the explanation of a compatibility constraint even if it takes several pages. I would remove repeated prose that neither changes a decision nor helps someone evaluate the work.

Matt Pocock's collection of engineering skills offers another example of packaging guidance into adaptable, composable pieces. I find that direction useful. Its stated design goals are not evidence that a collection of skills outperforms another workflow.

Keep the acceptance decision explicit

A small increment can still contain a bad assumption. Its advantage is that it gives us an earlier, more bounded opportunity to discover one.

The loop I would advocate is straightforward: define the next behavior, supply the context it needs, check the interpretation, inspect the change, collect evidence, and decide whether to accept it. Update durable guidance when a decision changes. If review becomes unmanageable, reduce the increment or improve the checks before asking the agent to produce more.

This approach has limits. It depends on the engineer's understanding, the quality of the evidence, and the cost of keeping guidance current. This article presents a working position, not a controlled study or a demonstrated productivity gain. Establishing such a gain would require comparable tasks and accounting for review, rework, defects, and maintenance—not just generation time.

I am comfortable delegating more of the work when I can explain what was requested, what changed, and why the evidence supports accepting it. That is the engineering standard I want to carry into AI-assisted development.

Which decisions in your workflow remain visible after the conversation ends—and what evidence tells you that the agent respected them?

Authorship note: The initial draft was generated with AI assistance from Armando Picón's arguments and a source review. Armando reviewed the article and approved publication. The Spanish translation was produced with AI assistance. The mobile example is hypothetical.