Monterro

Scandinavian Software Park by Monterro

Not just faster

How Agentic AI Changes What We Build

AI Builder Summit · Hanoi · 11 November 2026
Nils Hulth, Product & AI Lead, Monterro

AI makes software teams faster. The interesting part is how we use that power. This page collects the examples from the talk, with links to the original sources.

Three parts, nine areas

Better

The software itself

For every customer

Fit for each one

Further

What we know, and what we take on

Part 1 of 3

Better

The software itself

Quality

AI-written code does not have to be worse than code written by people. It can be better.

Linear

Jun 2026

Zero-bug policy since 2023: low-priority bugs are fixed within 7 days, high-priority bugs within 48 hours. An agent routes every bug report, and about 10% of bugs are fixed by an agent with no engineer involved. Giving the agent every bug did not work. Narrow bug types do: dead feature flags almost every time, failing background jobs about one time in three.

Security

An agent audits all the old code, and a second agent makes it prove every finding.

Performance

An agent finds slow database queries and proposes the fix.

Sentry

Nov 2026

An agent reads trace data from Sentry via MCP and ranks queries by how often and how slowly they run. It points to the file and line, and tests the SQL with EXPLAIN ANALYZE on a read-only copy. Then it opens a draft PR. Many fixes were one or two lines, such as a missing index.

Idea to try

Generate a year of realistic customers for your own product: bookings, invoices and history. Run the load test on that, and find the slow query before a customer does. An agent can write the generator from your schema. It can also learn how your real data is distributed without copying a single customer.

Reliability

When an alert fires, the agent starts at once.

Motorola Solutions

Jul 2026

Runs alarm systems for 911, the emergency number in the US. Downtime is a safety risk. An agent with read-only access to Grafana and Kubernetes, via MCP, reads metrics and cluster status and suggests a root cause. An engineer approves every change. Time to resolve an incident went from 45 minutes to under 4. Junior engineers get the same overview as senior engineers.

Part 2 of 3

For every customer

Fit for each one

Analysis

Problems that no customer reports are found, fixed and followed up.

PostHog

Jul 2026

AI goes through session recordings and finds where users get stuck. An agent drafts a fix as a pull request, and the team reviews and merges it. Later, an agent checks that the metric moved. If not, it reverts or reopens. In its first week on PostHog’s own product, it found two problems that no customer had reported. Both fixes were merged.

Ideas to try

  • Compare where users get stuck in the product with what the documentation says. Propose both the text change and the product change. Most teams already have the parts: session data or support tickets, the docs, and an agent.
  • Write one skill for each source of customer signals: support tickets, sales calls, NPS and app reviews, notes from customer success, interviews. Each skill turns its source into a report. An agent then reads all the reports and suggests ideas for the backlog.
  • Let an AI go through your customers’ configurations. Find features nobody uses, features used in ways nobody planned, and settings that lead to errors and incidents.
  • Let an AI go through customer data and find patterns you can use to develop the product.

Variation

Build several versions and let them compete before you choose.

Anthropic

Mar 2026

A generator agent builds a web page. A reviewer agent clicks through the live page with Playwright, takes screenshots and scores the design. They run 5–15 rounds, for up to 4 hours. In one run, the generator threw out a clean landing page and built a 3D museum room instead.

Custom

Each customer’s migration becomes a small program that an agent writes and checks.

Part 3 of 3

Further

What we know, and what we take on

Knowledge

Understand the product, not only the code.

Idea to try

Make internal videos from the system itself: the product, the codebase, the architecture and production. Make them again when the code, the architecture or production changes. They help with onboarding, and help everyone who works on a product they did not build.

Bigger scope

Agents made large rewrites affordable that teams had postponed for years. In all four cases, a large test suite made the change safe.

Ladybird

Feb 2026

C++ to Rust for the JavaScript engine’s front end: about 25,000 lines in two weeks, with zero regressions on 52,898 official JavaScript tests and the same bytecode. Rust had been rejected in 2024.

Idea to try

One scenario gives you the test, the docs, the training video, product marketing material and web content. Write a Gherkin scenario once. It runs as a test in CI on every commit. It rebuilds the help page for the flow every release. Playwright records a training video, narrated in each customer’s language. The same scenario can also produce feature sheets and demo scripts for sales and tech sales, and the feature page on the website with new screenshots.

Open source and tools to try

From the Monterro portfolio

Some of the examples in the talk came from Monterro portfolio companies.

Don’t just scale speed. Scale ambition.

AI gives product teams new capacity. The interesting question is what we spend it on.