AGENTS / 03

From developer tools to autonomous workflows.

Not “AI writes my code”. An engineering system where agents refine, implement, test and review — and humans approve the plan and the merge.

AGENTS / 01

End-to-end AI-agentic SDLC for Android

The pipeline I designed and run at Digidentity. A Jira ticket goes in; a reviewed merge request comes out.

IN USE · DIGIDENTITY · 2026CLAUDE CODEMCP · JIRA · FIGMA · GITLAB

01 AGENT
Refinement
Agent reads the Jira ticket and Figma designs and proposes a plan.
02 HUMAN
Human approval
An engineer approves or rejects the plan before any code is written.
03 AGENT
Development
Agent writes Kotlin / Compose code and tests against the approved plan.
04 AGENT
Testing
Tests run on Android emulators and physical devices on a dedicated server.
05 AGENT
Review
Agent reviews the change against the project's architecture and conventions.
06 HUMAN
Human approval
An engineer approves the merge request. Accountability stays human.

Runtime

A dedicated server with Android emulators and physical devices. Agents reach Jira, Figma and GitLab through MCP servers. Engineers approve the plan before coding and the merge request before it lands.

AGENTS / 02

MCP: controlled capabilities

Agents are only as useful as the systems they can reach — and only as safe as the way they reach them.

AGENT
MCP
TOOLS
SYSTEMS

Agents become useful when they can safely interact with real engineering systems. MCP gives them those capabilities through explicit, reviewable interfaces.

Connected via MCP

JiraTickets, acceptance criteria, status
FigmaDesigns and components
GitLabBranches and merge requests

Test environment

Android emulatorsInstrumented and UI tests
Physical devicesReal-hardware validation

Exploring next

Gradle
Logcat
Static analysis
Security scanning
Documentation

AGENTS / 03

Agent boundaries

Every agent needs boundaries. These are the ones I design for.

AGENT

CONTEXT
Agents work from the ticket, the designs and the codebase — nothing implicit.
PERMISSIONS
Least privilege: each tool exposes only the actions a stage needs.
TOOLS
Capabilities arrive through MCP servers, not ad-hoc scripts.
POLICIES
Architecture and coding conventions are part of the agent's instructions.
EVALUATIONS
Tests on emulators and real devices decide, not the agent's own claims.

EXECUTION
Agent acts within its permissions
VALIDATION
Deterministic checks: build, tests, devices
HUMAN APPROVAL
An engineer accepts or rejects
Also: auditability, traceability, rollback and explicit failure handling. This is what separates agentic engineering from vibe coding.

AGENTS / 04

Agent trace

What a run looks like, step by step. A simulated replay that mirrors the real stages.

SIMULATEDIllustrative replay · not real output

$ agent run "Add biometric authentication to sign-in"

Press Run to replay a pipeline run.

AGENTS / 05

Where this goes next

Specialised agents with their own tools and evaluations, still ending at a human.

DIRECTION · EXPLORINGNot in production. A model for where the pipeline could go.

HUMAN INTENT

PRODUCT AGENT
ARCHITECT AGENT
FEATURE AGENT
TEST AGENT
SECURITY AGENT
REVIEW AGENT
DOCS AGENT

MCP · TOOLSBUILD / TEST → SECURITY GATE → EVALUATIONHUMAN APPROVAL → RELEASE

End-to-end AI-agentic SDLC for Android

MCP: controlled capabilities

Agent boundaries

Agent trace

Where this goes next

LET’S BUILD SOMETHING DIFFICULT.