→ Engagement spotlight / Agentic engineering
A full manual regression stood between every code change and the customer. The suite in delivery is engineered to run it 24x faster
Three days by hand today, across around 300 test cases, six roles and several products. The target is under three hours, and it is an engineering target for work currently in delivery with no measured figure behind it yet.
It also widens coverage by three to five times across the full role and product matrix, replacing premium testing that was checked by sample.
→ The sequence / research to production
Six steps across three years.
01
Settle the argument first
A research sprint ran each candidate use case as a testable claim. Five of seven cleared against a target of three to five.
02
Move submissions off the legacy estate
Intake left the old platform for Salesforce across 2024 and 2025, with serverless AWS integrations behind it.
03
Encode the matrix as a specification
Roles, products and workflows written as a living specification the machine can read, so agents derive tests from it.
04
Shard the run
Playwright executes in parallel across role and product combinations. Each run seeds its own deterministic database.
05
Rebuild the workbench in slices
First slice is risk assessment and coverage: two tabs, roughly 60 fields, six to ten integrations.
06
Make the interface configuration
Build once, then configure per product and region, so onboarding a product stops being a development project.
Step 04 in detail / when a test fails, mint marks a human decision
↻ Loop A failure goes back into the specification. Fix the spec, regenerate the tests, run again.
→ The situation
A twenty-year-old underwriting workbench, and three days of manual testing before every release.
Underwriting in this business ran on a workbench between fifteen and twenty years old, built on Excel, VBA and SQL Server. Robotic processes reached into mainframes to fetch pricing and generate documents, every one of them manual. Fewer than a hundred people could use it at once, and the team wanted to extend the same book into the UK and Europe.
That single platform carried submissions and new business, risk assessment and coverage, rating and pricing, document generation, policy administration and the feeds into the financial ledger. Replacing it in one move was never a serious option, and the cost of leaving it alone was climbing every year.
Testing had a separate and more immediate problem. A full manual regression on the accident and health portal took three days, and it sat between every code change and the customer. Around three hundred test cases were run by hand across six roles and three to four products. Role-specific defects escaped, and premium recalculations were checked by sample, so the risk grew quietly with every change to a pricing rule.
The bottleneck was execution speed, the number of role and product combinations, and test data that would not stay still, none of which better test design could solve.
The division of labour
Engineers own the specification, and everything downstream of it is generated from that.
→ What was measured, and what is targeted
The numbers, and where they came from.
Full regression cycle, accident and health portal
Elapsed hours
Today, by hand~300 cases, six roles
Agentic suiteTarget, in delivery
The lower bar is under three hours, an engineering target for a suite currently in delivery, with no measured figure behind it yet. It is set out here because the plan behind it is specific and because the client has been buying against that plan for three years.
Use cases validated
The research sprint ran each candidate as a testable claim with a success threshold agreed before anyone wrote a prompt. Five cleared against a target of three to five. Measured.
Coverage expansion
Across the full role and product matrix, replacing premium testing done by sample. Projected for work in delivery.
Fields in the first slice
Risk assessment and coverage in a single embedded experience, with six to ten system integrations. The slice was sized so it could be finished and still prove the architecture.
Submissions intake left the legacy estate for Salesforce, the integrations run on serverless AWS services, and the delivery pipeline and operating model came with it. Delivered across 2024 and 2025.
Premium recalculation tests are generated from the pricing rules themselves, alongside button-state and value assertions for every role.
This is the control that makes the approach acceptable inside a regulated insurer. A named person signs off the specification before agents generate against it, and again before anything ships.
A generic form builder was on the table for the first milestone and we recommended deferring it. The wider form requirements were not clear enough to design against, and the risk assessment fields were already defined. It earns its place once more products are on the platform.
→ Why it matters
A research relationship became two production builds.
This started as a small research team embedded inside the client's own squads, with the client's developers holding the keyboard. Three years on, both delivery engagements are in Australia, and the relationship reaches well beyond the single regional team it began with. The research phase did its job, which was to settle whether the approach worked in this codebase, on this stack.
What makes a build of this size tractable is the division of labour. Engineers write the scaffolding and the specifications, agents generate and test against them at scale, and the specification is treated as the real deliverable. All of it runs on the Agent Centric Development Cycle, which concentrates the human effort at the front of the sequence and keeps a named person on the release gate.