Your Systems Assume a Human Did the Work

Zed just launched Delta, a standalone application built for a world where AI agents write the code, rather than bending the existing tools to fit. That is a remarkable thing for a developer-tools company to conclude: that decades of version control carry an assumption no longer true. Changes used to arrive in human-sized batches, at human speed, from a person who could explain each one. Coding is only where that assumption broke first. The same one sits quietly underneath how your business reviews everything.

The assumption nobody wrote down

Think about what approving something actually assumes. Signing off a contract assumes someone drafted it and can tell you why clause four says what it says. Approving a report assumes a person made the judgment calls buried in it. Reviewing a campaign assumes there is an author to question. None of that was ever a rule anyone documented, because it did not need to be; work came from people, and people can be asked. As AI drafts more of the raw material, that assumption fails silently. The work still arrives, still looks professional, and still gets approved.

Three things that break

In practice the failure shows up in a consistent order, and none of it announces itself.

What breaks	How it shows up
Volume overwhelms review	Same process, ten times the output, shallower checks
Attribution disappears	Nobody knows which parts were generated
The explanation gap	The owner cannot say why it says what it says

The first row is the one that quietly does the most damage, and it is the same dynamic behind reviewers who end up rubber-stamping. A process built for five documents a week does not fail loudly at fifty. It just stops being a real check.

Record origin while it is still free

The single highest-value habit is also the dullest: note what was AI-generated, what a person wrote, and what a person changed. Do it as routine, not as a special exercise, because the whole point is that the trail exists before anyone needs it. You cannot reconstruct this later. Six months after the fact, nobody can tell you which paragraphs of a proposal were drafted by a model and which were considered line by line, and that is exactly when someone asks. The same logic makes a change log the core control for agents that rewrite their own instructions: traceability is cheap in advance and impossible in hindsight.

Stop reviewing everything equally

If AI has increased your output tenfold and your review process has not changed, you are not reviewing more carefully, you are reviewing everything worse. The fix is to differentiate. An internal draft needs a glance. A document going to a client, touching money, entering a permanent record, or informing a decision you would have to defend gets real scrutiny from someone who understands it. Deciding that split explicitly, and writing it down, is what keeps quality from silently eroding as volume grows, and it is how you avoid the polished-but-wrong output we called workslop.

Somebody has to be able to explain it

Here is the test that cuts through everything else: for any significant piece of work your business produces, can a named person explain, in their own words, why it says what it says? Not who ran the prompt. Why the content is right. If the answer is no, you have output without accountability, and that gap is where the real risk lives, whether the questioner turns out to be a client, an auditor, or a regulator. It is the same accountability principle behind governing the agents you deployed, applied to the work rather than the tools.

Cheap now, painful later

None of this needs to be a project. It is a few habits adopted early, at a point when they cost almost nothing. The reason to start before it feels urgent is that every part of this is retroactive-hostile: you cannot add attribution to work already produced, and you cannot manufacture an explanation for a decision nobody recorded. Zed rebuilt a category of tool because the old assumptions stopped holding. You do not need new tools. You need to notice that the same assumption is sitting in your process, and adjust it while adjusting is easy.

Frequently Asked Questions

What happened?

Zed launched Delta, a standalone application built specifically for a world where AI agents write the code, rather than adapting the existing tools designed for human authors. That is a striking admission from people who build developer tools: the assumptions baked into decades-old version control, that changes arrive in human-sized batches, at human speed, with a person who can explain each one, no longer hold when an agent produces a hundred changes in an afternoon. Coding is simply where the strain showed first. The same assumption sits underneath how most businesses review any work.

Why does this matter outside software?

Because every review process you have was designed around a person. Approving a contract assumes someone drafted it and can explain their reasoning. Signing off a report assumes a human made the judgment calls inside it. Checking a marketing campaign assumes there is an author to ask. Those assumptions are quietly failing as AI drafts more of the raw material. Nothing breaks loudly; the work still arrives and still looks fine. What changes is that the person approving it can no longer reconstruct how it came to say what it says.

What actually goes wrong?

Three things, in order of how often we see them. Volume overwhelms review, because a process built for five documents a week silently degrades when it receives fifty. Attribution disappears, so when something turns out to be wrong nobody can say which parts were generated, which were edited, and on what basis. And the explanation gap opens: the person who owns the output cannot answer why it says what it says, which is fine until a client, an auditor, or a regulator asks. None of these are AI failures. They are review processes meeting a volume and origin they were never designed for.

What should we change?

Three practical habits. Record origin as a matter of routine, noting what was AI-generated, what a human wrote, and what a human changed, so the trail exists before anyone needs it. Match review depth to consequence rather than reviewing everything equally, because uniform review across a tenfold increase in volume just means shallow review everywhere. And keep a named owner who can explain each significant output in their own words, which is the real test of whether your process still works. If nobody can explain it, you have output but not accountability.

Is this urgent for a small business?

It becomes urgent the moment AI is producing work that leaves your building, goes to a client, or informs a decision you would have to defend. Before that, it is cheap housekeeping. The reason to do it early is that retrofitting attribution is painful: you cannot reconstruct six months later which parts of a document were generated and which were considered. Starting the habit now costs almost nothing and means that the first time someone asks a hard question about a piece of work, you have an answer rather than a shrug.