Loading
MA
hello@byanees.comDownload CV
← All work

Agentic AI · Systems Limited, 2026 – PresentSupervised Agentic Delivery

Three mode agents, picked by the work item. The agent does everything I can undo; merges, migrations, and database writes stay with me. Runs in Cursor, Claude Code, or any other agent.

Company

Systems Limited

Role

Product Engineer

Period

2026 – Present

Stack

Cursor, MCP

3mode agents, picked by the work item, in place of an orchestrator
5tools over MCP: Codegraph, Context7, Figma, Bytebase, Azure DevOps
0merges, migrations, or database writes left to the agent
how it workshow it works

The work item picks the mode. I hold the gate

Pick a step to replay it

agent · everything I can undoWork itemAzure DevOps · TeamsL3 modebug · hotfix · verifyFeature modeuser storyTriage modeno work item yetPR → releaseside effects checkedPR → developafter gap analysisDraft bug WIroot cause, no codeMemerge · migrate1prod bug24review3confirmrules load in every chat · 5 tools over MCP

The problem

During hypercare, from September to October 2026, I worked on 40+ production incidents under an 8-hour SLA. Fast fixes only count if nothing else breaks. The Cursor setup I had, long specialist agents under an orchestrator, loaded a lot into every chat, and its guardrails held only as well as that day's prompt.

What I built

  1. Three mode agents, picked by the Azure DevOps work item: a prod bug, hotfix, or test verification goes to L3 mode, a user story to feature mode, and a Teams message with no work item yet to triage mode
  2. L3 mode reads the work item and every comment on it, reproduces the issue on a test host, writes the smallest fix it can, then checks what else the fix could touch (notifications, scheduled jobs, integrations, config) before opening a PR to the latest release branch
  3. Feature mode runs a gap analysis before any code and opens its PR to develop. Triage mode finds the root cause and drafts a bug work item, writes no code unless I ask, and hands off to L3 once I confirm
  4. Six skills as playbooks: bug fix, feature delivery, triage, GitFlow, PR and work item write-ups, and live library docs
  5. Five tools over MCP: Codegraph first for the codebase, Context7 for library docs, Figma, Bytebase, and Azure DevOps
  6. Rule files that load in every chat: the database is read-only by default, topic branches only, no merges, no migrations, a browser on four test hosts and nowhere else, and every work item comment stays a draft until I approve it
  7. It replaced the orchestrator and its specialist agents, which are now archived
  8. Nothing in it is tied to Cursor: the modes, skills, rules, and MCP tools are plain files in the repo, so the same setup runs in Claude Code or any agent that loads rules and calls tools over MCP

StackCursor, MCP, LLM agents, Azure DevOps, Context7, .NET 9 and Angular

next case studynext case study

Request to Pay over App and USSD

PaymentsDPL, 2024 – 2026

Request to Pay over App and USSD

Led delivery of Request to Pay: a merchant requests money from a customer, who can pay from a push notification or a USSD prompt, and can never pay twice.

4,000+merchants adopted Request to Pay

Built with.NET, Microservices, USSD and Push notifications

Read case study

Building something that has to work on the first attempt?

Open to product, full stack & backend roles.

Get in touch
Download CVDownload CV
Let's talk