How We Build Software
Agentic AI Development
Agentic AI works inside the workflow, not beside it.
We build secure software for organizations that can’t afford to get it wrong. AI agents do real work across the whole lifecycle, under human control at every step. Here’s how.
How agentic AI fits into the way we work
A few ideas hold the whole approach together. They’re why work produced by AI agents meets exactly the same standard as everything else we ship for you.
01
Agentic AI works inside the workflow, not beside it
The agents use the same artifacts we use every day: issues, milestones, the board, pull requests, reviews. There is no parallel AI track to reconcile later.
02
Humans gate, agents create
Nothing reaches the codebase, and no work is created, without explicit human approval. Gates are hard stops, and silence never counts as approval.
03
Everything traces to intent
Plans define objectives; features trace to objectives; tasks nest under features. Before implementing, the agents read the plan the work belongs to.
04
One source of truth
A single internal toolchain defines how the agents behave on every engagement. The process is consistent and proven, never reinvented project to project.
05
Quality is enforced, not requested
Test-first development, code review, and security review run automatically inside delivery, never as optional steps that get skipped under deadline pressure.
06
The system learns
Every project feeds back: learnings become improvements to skills, standards, and architecture decisions. The way we build gets sharper with each engagement.
Agentic Workflow From Requirements to the Released Software
Five named steps run every engagement, with AI agents doing the work inside each step and a human owning every gate. The first two are written, committed documents we create with the client; the last three run as tracked work in GitHub or Jira, where committed intent becomes epics and tickets and every change traces back across that line. Where the merged code goes next, whether a release, an app store, or a cloud environment, depends on the client’s system and follows the client’s delivery process from there.
VERSION-CONTROLLED DOCUMENTS · PM-owned
01
Brief
The client’s intent captured and stress-tested into a committed brief: objectives, scope, and target users. We propose, the client confirms, never a blank left to fill in later.
02
Spec
The committed brief translated into a build-ready requirements spec: user stories, acceptance criteria, and technical design. Locked against silent overwrites, not against change.
TICKETING SYSTEM · GITHUB / JIRA · Engineering
03
Plan
The confirmed spec broken into objectives, features, and tasks, sequenced, with dependencies mapped. Where documents become tickets.
04
Code
Agents implement each task against its acceptance criteria, test-first, and open a focused pull request per task. Small, reviewable changes with a clean, conventional history.
05
Review
A human-gated review blends automated code and security review with a human verdict: approve, request changes, or block.
Humans gate, AI creates. Nothing is written to the codebase, and no work is completed, without explicit human approval. Silence never counts as approval.
Everything Traces Back
Merged code
Task
Feature
Objective
Plan
Spec
Brief
Before implementing anything, the agent reads the plan the work belongs to. The bigger picture is required context, never optional, so every line of code traces to a reason for existing.
Industry-leading AI, our standards on top
Frontier AI models
Claude and OpenAI, in parallel
Our AI workflow uses both Anthropic’s Claude and OpenAI’s models together, among the most capable AI coding models available, run in parallel on the same task alongside the same git, build, and review tooling our engineers use every day.
The A-CX layer
Integrated into Codex and Claude
On top of both sits our own layer of skills, automated agents, and quality rules, installed directly into the Codex CLI and Claude desktop/CLI. We distribute and update it through the marketplace, so every developer always works with the latest A-CX standards.
The right model for each task
We assign a deliberate AI model and effort level to each task, so routine work runs light while planning, architecture, and security judgment stay on the strongest model. The result is sound engineering decisions without wasted effor.
Fast
Lightweight tasks
Background work, monitoring, simple transforms, and pattern detection. Quick, low-cost, high-volume.
Balanced
Everyday delivery
Code review, security review, implementation, and standard development tasks.
Reasoning
High-judgment work
Architecture, planning, requirements analysis, and security judgment, where getting it right matters most.
When the output isn’t good enough, we escalate using a stronger model. Planning, architecture, and security judgment run on the most capable model by rule: a bad decision early costs far more than the difference in compute.
Quality is enforced, not requested
Our quality standards run automatically inside delivery, never a checklist someone might skip. Exceptions are explicit, documented decisions, never quiet omission
Test-first
New code is written against tests first. Behavior is specified before it’s built, not validated after.
Security review
Security review runs automatically on every feature delivery, designed in, not bolted on afterward.
Code review
Every change is reviewed and approved, with automated checks green, before it merges.
Clear history
A disciplined, conventional commit history keeps the codebase auditable and easy to hand over.