AI agents for the development lifecycle
An agentic spec-driven development (SDD) system for GitHub Copilot, with custom agents and skills that support development teams from discovery to QA.
-
DiscoveryExplores ideas, writes and refines Jira stories
-
DevelopmentSpec, plan, tasks and implementation (SDD)
-
QAGenerates test cases and runs them
-
DeploymentNext stagecoming soon
human review at every stage
This is a summary of my current work at Banco Bolivariano. For confidentiality reasons, it leaves out internal implementation details.
The challenge
Coding assistants are good at writing individual functions, but a complex project starts long before the code and ends long after it. Requirements have to be understood, user stories written clearly, solutions designed, implemented, tested and deployed, all within an environment with strict security and governance rules.
What I built
An agentic spec-driven development (SDD) system that runs on top of GitHub Copilot. Each stage of the lifecycle has its own agents and skills:
- Discovery: agents that help explore ideas for new user stories, draft them, refine them and publish them directly to Jira.
- Development: the full SDD workflow. The specification is the single source of truth, and the technical plan, tasks and implementation are all derived from it.
- QA: agents that generate test cases from the specification and run them.
- Deployment: the next stage to cover, still being planned.
To connect the agents to the real world, I orchestrated existing MCP servers for Jira, Figma and the bank’s internal systems, so the agents work with up-to-date information instead of assumptions.
Governance
Bringing agents into a bank requires clear rules, so I also lead the operating and governance model for SDD-based development. Two principles are non-negotiable:
- Mandatory human review at every stage: agents propose, people decide.
- A tool-agnostic model, so it keeps working if the assistant or the workflow changes.
The next step is living specifications: keeping the spec up to date alongside the code instead of letting it go stale.
Results
The 3 teams already using the system doubled their productivity (2x), measured through throughput, cycle time and velocity. The current goal is to sustain that level and roll the system out to every development team in the bank.
Stack
GitHub Copilot, custom agents and skills, MCP servers, Jira and Figma. Outside the bank, I also work with Claude Code and with Spec Kit, for which I published a community preset.