Skip to content
Stop Token

Blog · 24 October 2026 · Alex Barakov

Product builders: no results yet.

In 2026 tech companies are piloting a new role: the product builder. One person comes up with a feature, designs it and builds it with AI agents, work that used to take a product manager, a designer and several engineers. Five companies have said in articles, podcasts and other open sources what came of it: more code, little of it reaching users, and nobody has measured the effect on the business yet.

The research as a PDF, 9 pages →

Meme: “We're product builders now.” “So you ship more features, or ship them faster, right?” Silence. “…right?”

The point

+220%

code changes in Meta's internal systems in a year

+36%

changes that users actually saw

Meta internal data, published by Reuters, 2026

  • Some developers, product managers and analysts become builders. The other roles become their support: data, review, hard cases, the platform.
  • Delivery really is speeding up: code gets written faster, and everyone notices. Quality and reliability lag behind: rollbacks, outages and tech debt grow. Discovery and analytics are not speeding up yet, and teams end up with empty backlogs. Analytics gets done partly or skipped.
  • Nobody has measured the effect on the business.

Five pilots

Who tried it and what came out

Only companies that described how they set up the pilot and what came out of it are listed here. LinkedIn, Coinbase, Block and Ramp announced programmes but showed no results. OpenAI's Harness team simply grew from 3 to 7 engineers, which is not a builder pilot. So none of them are here.

Before

A team of 10–20: engineers, a product manager, a designer, an analyst, a researcher, a data engineer

After

A group of 3–5: 3–4 builders, usually former engineers, and one person who decides what to build. Designers and analysts sit in a shared team across groups

Meta's internal slide, published by Reuters

  • Meta
    Partly rolled back

    What they did

    • Groups of 3–5 instead of teams of 10–20
    • ~1,000 people from Reality Labs, the VR and AR division

    What came out

    • Code +220%, changes for users +36%
    • Major outages +40%

    Source ↗

  • Atlassian
    Rolled back

    What they did

    • ~10 projects of selected builders
    • Product managers, designers and engineers in one role

    What came out

    • Fast first weeks, then a sharp slowdown
    • No figures published

    Source ↗

  • Intercom
    Outcome unknown

    What they did

    • ~500 people in R&D
    • 94% of code changes created by an agent, designers ship their own fixes

    What came out

    • Three times the changes per person in 16 months
    • Bug queue −54%, idea to release −39%

    Source ↗

  • epilot
    Continuing with limits

    What they did

    • A German B2B software company. For 60 days, 50+ people outside engineering got access to code

    What came out

    • 40 of 629 changes came from 11 people outside engineering
    • One change reached production past the automatic checks

    Source ↗

  • Uber
    Outcome unknown

    What they did

    • A year of product managers prototyping with AI

    What came out

    • “Two hours of prototyping replaced four weeks of discussion”
    • Prototypes get taken for the finished solution

    Source ↗

“Outcome unknown” means the company has not said whether it met its goal.

What breaks

Four problems that keep coming back

More code, same value

+30%

code after AI agents, Harvard, 718 companies

  • Finished tasks did not grow
  • At Meta changes for users +36% against code +220%

Nobody decides what to build

1 : 30–40

one product manager or designer per 30–40 engineers, as Atlassian's teams felt it

  • Everyone went off to write code, nobody set the direction
  • It used to be 1 to 10. The pilot was rolled back

Review cannot keep up

+40%

major outages, Meta

  • Firefighting time +70%
  • Intercom: code review became the bottleneck, quality dipped

People burn out

74% → 55%

satisfied with work, Meta

  • Former managers asked to come back to management
Expanding brain meme, self-service agentic analytics for builders: ask an analyst; ask an AI agent; file a ticket for an analyst to fact-check the AI agent; just skip the analytics step
I see the fourth option more often than I would like.

My observations

Analytics falls behind

Features ship faster than anyone checks whether they worked. I see three reasons: builders lack the skill, they would rather not dig in themselves, and the agent has nothing to stand on.

  1. Analytics gets done halfway

    A builder ships a feature and does not check whether it helped. When they run an A/B test, they often cannot set it up or read it properly.

    In open sources

    • Bookingnot a pilot

      80% of tests never checked in advance whether there were enough users to see an effect. Courses did not help; hints in the testing platform did

    • Uber

      Making a prototype became easy; working out what it tells you did not

  2. Builders don't want to do it

    Builders do not want to dig into data themselves and go to an analyst. The load on analysts does not drop.

    In open sources

    • Intercom

      The company president used to file a request with an analyst and wait. They built an agent that knows the company's data

    • Rampnot a pilot

      1,476 questions to the agent in 4 weeks, 66 in the data help chat

  3. The agent has nothing to stand on

    Without instructions and shared metric definitions the agent writes queries from scratch, numbers disagree, and the builder goes back to an analyst.

    In open sources

    • Anthropicnot a pilot

      Without instructions the agent was right at most 21% of the time, with them over 95%. Without upkeep, 65% a month later

    • Meta

      Moved its agent to ready, checked queries and bought a startup that builds a shared metrics layer

Code and interfaces

  1. Code gets sloppy, debt piles up

    Tests, docs and clean-up are skipped for speed. Six months to a year later, every change takes longer and costs more.

    In open sources

    • Intercom

      When AI started writing most of the code, code quality fell, then they pulled it back up

    • Atlassian

      Fast first weeks, then a sharp slowdown

  2. Products look alike

    Agents build interfaces from the same templates. Nobody thinks about whether the product is easy to use.

    In open sources

    • Uber

      Teams cling to the first version of a prototype

    • Intercom

      Taught agents its own set of interface elements: code using it went from 7% to 96%, yet a quarter of screens still have errors

Roles

Who moves where

My generalised scenario from the pilots. Some developers, product managers and analysts become builders. Every role becomes support for builders. No company has gone all the way.

What changes in each role

Developer
Writing codeGiving the agent tasks and checking its code. Plus part of the product job (what to build and why) and of the analyst's (checking whether a feature worked)
Product
Mock-ups, simple featuresDeciding what to build and in what order, talking to customers. Plus part of the developer's job (building a feature with the agent) and of the analyst's (running a simple test, reading a metric)
Analyst (DA, DS)
SQL, routine pulls, simple A/B testsComplex tests and research, teaching other roles their skills
BI/AE/DE
Building metrics, data marts, dashboards, pipelines from templatesRunning the domain knowledge base and shared metric definitions for the agent (semantic layer), checking its answers (evals), owning which data is right, keeping an eye on what builders make

Builders team

ships features with agents

Who

Some developers, product managers and analysts

What they do

Come up with a feature, design it and build it. The agent writes the code, queries and prototypes

Support for builders

enablement

Who

Every role: BI, AE and DE, designers, and the developers, product managers and analysts who did not become builders

What they do

Keep builders from breaking things: data and metrics the agent can trust, checks of its answers, the platform, a shared set of interface elements, training

People

Who becomes a builder

My observation
Product managers and analysts are easier to retrain: they learn to code with an agent quickly. Thinking about the business and making decisions are harder to teach.

Other views

  • Intercom: the best engineers will take on part of the product job ↗
  • Atlassian: only a few product managers and designers will become builders ↗
  • Harvard and P&G: with AI, people from R&D and from commercial propose more balanced solutions; without it each pulls their own way ↗

What happens to newcomers

If you run a pilot

What to measure

  1. Artefacts: is there an effect

    Look at what builders made: features, tests, reports. Did it reach users and move a metric.

  2. Logs: are there attempts

    Who tries to do the task themselves with the agent, how often, and at which step they give up and go to a specialist.

  3. Barriers: what fails in the agents

    Why the job someone came to the agent for (JTBD) was not done: no data or access, a wrong answer, a missing skill.

If you run a builder pilot and measure its effect, tell me how.

The research as a PDF, 9 pages →

Alex Barakov

Author

Alex Barakov

Comments and objections are welcome on LinkedIn or at info@stop-token.com.