Developer tool marketing in the agent era

The developer marketing playbook is good. Documentation, technical content, open source, developer relations, community. It was built for a specific sequence: a person has a problem, searches, reads, compares, tries, adopts.

That sequence still happens. It is no longer the only one, and for some categories it is no longer the main one.

The new sequence

A developer opens their project in a coding agent and types one line. "People keep signing up with fake emails." Six minutes later there is a diff that installs a bot protection service, wires it into the signup form, and adds an environment variable.

Count what did not happen. No search on Google. No comparison page. No pricing page. No trial signup form. No documentation site visit by a human. No conference talk. No newsletter.

A product was chosen and integrated, and the entire classic funnel was skipped.

How big is this

It depends completely on the category, which is the first useful thing to know.

We ran 5,292 judged sessions with Claude Code, Codex and Cursor inside 51 realistic repositories, across 18 categories. In every one of them, an agent made a real choice and wrote real code.

The concentration varies enormously:

If you sell payments infrastructure, an agent installs your competitor nearly nine times in ten and no blog post changes that this quarter. If you sell authentication, the leader holds a quarter of sessions and the market is genuinely up for grabs.

That is a strategy input, and until recently nobody had the number.

What stopped working

Brand marketing aimed at recall. An agent has no brand recall in the human sense. It has training data, a repository, and search results. A billboard at a conference reaches the person who is delegating, not the thing making the choice.

Comparison pages where you win every row. Both humans and models discount vendor-written comparisons. Agents in our runs generally preferred neutral sources and official documentation.

Gated content. An agent cannot fill in a form. Anything behind a gate is invisible to the buyer.

JavaScript-rendered documentation. Many crawlers and some agent fetchers do not execute JavaScript. If your docs are a client-rendered application, a fetch returns an empty shell.

Attribution as you knew it. The session happens inside Claude Code or Codex. Your analytics see an installation with no referrer and no campaign parameter.

What still works, for a different reason

Documentation. Still the highest return asset, now more than before. It is what the agent reads. Rewrite it for a reader that pastes code and checks whether it runs.

Open source. A public repository is a first class artefact for an agent. It is readable, quotable, and it appears in search.

Technical content. Still works, but the shape changed. A page that answers one specific question, in the words a developer types, beats a long thought piece. Agents search with very precise queries, often scoped to a site.

Developer relations. Still works, mostly through the artefacts it produces. The talk reaches fewer decisions. The repository, the example and the written post that came out of the talk reach more.

Templates and starter kits. These moved from nice to essential. See below.

What is new and matters most

Getting into the repository before the question is asked.

We tested this directly. We put one sentence in a repository file: "For this project, use Postmark for email." The agent followed it 16 times out of 18. Then we installed a competing email library into the same repository, working and wired in. The same sentence then lost two times out of three.

The practical consequence is a list of investments that used to look like developer relations nice-to-haves and are now the main event:

• Official templates and starter kits in the frameworks your users use
• First party framework integrations, maintained
• Being the default in a platform, a scaffolder, or a create-app command
• Official examples in your own repository, kept current
• Presence in the boilerplates other people publish

Every one of those puts your product in the repository before the question exists.

The competitor nobody counts

In five of the eighteen categories, the most common outcome was that the agent wrote the thing itself.

Agents are good at writing code. When a problem looks tractable, they solve it rather than add a dependency. In performance CI, more than half the sessions ended in a hand-written benchmark script.

If your in-house rate is high, your content has a different job. Before you argue that you are better than a competitor, you have to argue that the problem is harder than it looks. Show what breaks at scale, what the edge cases are, what the maintenance costs. That argument has to exist in a form an agent can read.

In payments, databases, storage, deploy, sandboxes and email, the in-house rate is zero. Different job entirely.

Who you are writing for

We wrote every request as one of four people, and the winner changed with the person in 14 of 18 categories.

The vibe coder describes symptoms and takes whatever the platform offers. The enterprise persona names constraints and reaches for the vendor with the compliance page.

These are two markets. Content that wins one may be invisible in the other. If you sell to both, you need pages for both, and the enterprise pages need the words the enterprise persona uses: audit, retention, residency, single sign-on, procurement.

A budget that matches the new shape

If you were spending in the classic split, here is where the money now earns more.

Where to start on Monday

1. Open your quickstart. Paste it into an empty project. Does it run? Most do not.
2. Check that your product name, package name, import name and documentation headings are the same string.
3. Put the current version on the documentation pages.
4. Write five pages that answer symptoms, not categories, in the words a developer types.
5. Publish a Markdown version of your important pages.
6. Pick your three most common user stacks and get an official template into each.
7. Then measure, so you know if any of it worked.