Building a Structured AI Agent-Driven Engineering Setup for Travly

Introduction

Travly covers hotel search, booking, subscriptions, and trip management for everyday travellers. The engineering team consists of five people: a backend lead, a frontend lead, an apps lead for React Native, a QA engineer, and a PM. By late 2025, the team was using four different AI tools: Cursor, Copilot, ChatGPT, and Gemini. Individual engineers were working faster. The team as a whole was not, because knowledge from one session never carried over to the next person or the next day.

The product roadmap made this gap impossible to manage. The hotel pipeline migration, Booking API rollout, Pro Redemption flow, Trips and Wishlists rebuild, and discovery surfaces overhaul were all in progress at the same time. When Vio disconnected its API, the team also had to build an in-house price comparison module from scratch. Travly needed a partner who could change how the team operated, not simply introduce another tool.

The Challenge

The core problem was that everything an AI needed to be useful existed only in engineers' heads. There was no documented record of which database queries hit the read replica, which pricing branches were tax-sensitive, or which supplier integrations carried risks that did not appear in the codebase surface. Every AI session began with no prior context. This led to a set of problems that built on each other:

• Four AI tools across the team with no shared context between sessions
• Architecture decisions and naming conventions communicated verbally, never recorded in a way any tool could use
• Production errors in Sentry were accumulating with no automated path for diagnosis or resolution
• Each product spec took half a day to write, requiring several rounds of meetings before work could begin
• Separate specialization lanes for backend, frontend, and mobile that created bottlenecks whenever a feature touched more than one area
• No reliable way to track delivery beyond standup updates and searching through Slack history

The Approach

In early February 2026, we moved the entire team onto Claude Code on Anthropic's MAX plan. The MAX plan uses flat-rate pricing, which makes long-running, multi-step agent workflows practical in a way that per-message pricing does not. Before any feature work resumed, we spent February and the first half of March making deliberate architectural decisions about how the system should work, who should own what, and what the team needed in place before agents could be trusted to run unsupervised.

Turning Convention Knowledge into Shared Infrastructure

Our first decision was to get the team's knowledge out of engineers' heads and into a form every tool could read. A CLAUDE.md file was added to every repository, documenting working conventions: architecture decisions, naming patterns, database routing logic, pricing edge cases, and known risks in supplier integrations. 

We used Claude to produce the first draft by reading the codebase directly, then had the relevant lead fill in what could not be inferred from the code alone. This file is what every agent reads before working in the repo, and what every new engineer reads on their first day.

We structured the agent architecture around five agents in-repo, each with a defined scope and a human owner to ensure accountability:

On the backend, each agent runs in its own isolated environment using worktree-test-env.sh, so all five can operate at the same time without interfering with each other's processes or database connections.

A Skills Library for Recurring Work

Rather than letting agents improvise on every task, we built a skills directory in each repository: a set of saved, reusable instructions that encode how specific types of problems should be handled. The backend repository has 17 skills organised into four categories:

• Error diagnosis: sentry-issue-resolver, sentry-repro-verifier
• Database and query work: db-explorer, cache-advisor
• Infrastructure: aws-explorer, provider-hardener, api-version-manager
• Workflow: jira-ticket-creator, nest-test-generator, security-scanner, pr-review-checklist, logger-upgrader, pr-review, nest-module-scaffolder, service-refactorer, e2e-test-manager, flyway-migration-generator

The frontend and mobile repositories each have their own skill sets matched to their tech stacks and common failure points.

Redesigning a Product Spec Pipeline

We redesigned how product specs were produced. The PM took ownership of a seven-agent pipeline: Intake, Scope, UX Flow, Tech Plan, Delivery, Critic, and Analytics. 

A product spec that previously required half a day and multiple meetings now takes about 15 minutes to produce. We made the Critic stage non-optional: its job is to catch specs that are too vague and would result in an implementation that is technically correct but does not do what was intended.

Connecting Tools Through MCP

We connected all key tool integrations through MCP: Slack, Jira, GitHub, Sentry, Crashlytics, Mixpanel, Figma, AWS, Playwright, and Singular. This gave agents access to live information rather than static data. 

We also built a standalone delivery tracker that scans relevant Slack channels and DMs and publishes a daily summary of what had shipped, what was in progress, and what was blocked, with the responsible person named against each item. It took over from the daily standup as the team's main source of delivery information.

The Outcome

The system went live on March 17, 2026. The results over the next 60 days were measurable across every dimension of delivery.

Before Arbisoft, the team was spending half a day writing each product spec, 1 to 3 days moving a single ticket to a reviewed pull request, and 8 to 12 weeks integrating a new hotel supplier. Production bugs were resolved manually. Test authorship fell entirely on engineers. Four AI tools were in use but none of them shared context, so every session started from zero.

After we built the system, PRD authoring time dropped by over 90%, supplier integration dropped by roughly 90%, and the team tripled its sprint output without adding a single person to the team.

When the hotel data pipeline entered an 11+ hour failure loop in April, peak memory at 3,143 MB, 57 database connections, and Aurora CPU at 94%, three diagnostic skills ran in parallel: aws-explorer for CloudWatch metrics and pod logs, db-explorer for failed sync records, and sql-query-optimizer for the query spike. Two days later: no crashes, memory at 901 MB (down 71%), database connections at 38 (down 33%).

In a single frontend session, 63 analytics events were instrumented across 48 modified files and 5 new files, covering 4 conversion funnels: auth, search, hotel, checkout, subscriptions, and trips. Verified through Playwright. Confirmed in Mixpanel Live View. The Test Engineer agent wrote 322 backend tests. The end-to-end suite runs 194 checks in 32 minutes. Frontend build times dropped from 19 minutes to 4 to 5 minutes on Blacksmith.

Engineering capacity stopped being the constraint. The team's focus moved to spec quality, prioritisation, and judgment.

Value Delivered

The numbers above reflect more than process improvements. Each one represents real time, effort, and cost returned to the team.

• PRD authoring dropped from half a day to 15 minutes per epic, over a 90% reduction in time spent before any engineering work could begin
• Hotel supplier integration went from 8 to 12 weeks to 5 to 6 days, roughly 90% faster
• Sprint output tripled with no additional headcount, meaning the same team budget delivered 3x the engineering output
• 80 production bugs were resolved autonomously, with zero engineer hours spent on diagnosis, investigation, or fix
• 9 tickets moved from open to reviewed PRs in 58 minutes, work that would previously have taken 9 to 27 days of engineer time
• 322 backend tests were written by the Test Engineer agent, test authorship that would otherwise have fallen on engineers
• 60,000+ Sentry errors were cleared in a single automated pass, eliminating a backlog that had been accumulating for months without anyone touching it

Arbisoft did not replace the team. It changed what the team was capable of. The same five people, with the right system in place, delivered at a scale that the previous model could not sustain.

Technology Used

Claude Code on Anthropic's MAX plan was used as the single AI coding tool across all three repositories. It was connected to Slack, Jira, GitHub, Sentry, Crashlytics, Mixpanel, Figma, AWS, Playwright, and Singular through MCP.

• Backend: Aurora (PostgreSQL), Redis, Typesense, Flyway
• Frontend: Next.js, Zustand, TanStack Query, Jest, React Testing Library, Playwright
• Mobile: React Native (iOS and Android), Detox, Maestro