Forge

How we use AI to deliver software

Forge is the system behind all our work. AI agents handle analysis, building and testing, each task goes to the model best suited to it, and a senior person approves every stage.

Forge architecture A senior human lead directs the Forge orchestrator, which coordinates five agent roles. Agents call several AI models through a model router, and outputs pass human approval gates before reaching your repository and cloud. Senior leadsets goalsapproves Forgeorchestratortrace · gates · audit Analyst agentsArchitect agentsBuilder agentsTester agentsAuditor agents ClaudeGPTGeminiLlama / Mistral Model router Agent teams Humangate↓ yourcloud

How it works

Five stages, each with a named owner

  1. 01

    Understand

    Analyst agents + senior lead

    Agents read your policies, specifications and existing systems, then draft user needs and requirements in hours. Our lead checks every one.

    Requirements · Risks · Assumptions

  2. 02

    Design

    Architect agents + senior lead

    Agents propose designs that follow the Technology Code of Practice. A senior architect chooses one, adapts it and signs it off.

    Design decisions · Threat model

  3. 03

    Build

    Builder agents + senior lead

    Agents write code and content in small changes. A senior reviewer approves each change before it ships.

    Code changes · Infrastructure as code

  4. 04

    Verify

    Tester agents + senior lead

    Agents test for bugs, accessibility problems and security gaps, and check AI features for accuracy and bias.

    Test results · Accessibility report · AI evaluation

  5. 05

    Release

    Release agents + senior lead

    A person approves each release to your cloud, and any release can be rolled back. Documentation is written as we go.

    Release notes · Runbooks · Transparency record

Safeguards

Speed only helps if you can trust the result

We built Forge for organisations that answer to auditors, regulators and boards.

Approval gates

Nothing reaches production until a named person approves it.

Labelled AI work

We record which model produced each piece of work and who reviewed it.

Requirements linked to tests

Each requirement links to the code that meets it and the tests that check it.

Accessibility checks

Automated WCAG 2.2 checks run with every set of tests. A failure stops the release.

Security checks

We scan dependencies, secrets and code, and test AI features for prompt injection and data leaks.

Data stays where agreed

Client data stays in the regions and cloud accounts we agree with you.

Forge Guard

Use the same controls on any AI agent.

We are packaging these safeguards as Forge Guard, so security teams can assess and control agents built by anyone.

How Forge Guard works

Choosing models

The right model for each task, and the freedom to change.

We test models on each type of task and use the one that does it best: the most capable models for design and complex code, faster ones for tests and documents, and open models on your own infrastructure when data must stay there.

  • Prompts and tests work across providers
  • Model changes are measured and reversible
  • You benefit as model prices fall

What you receive

Working code and the evidence behind it

Working, tested software in your repository
Requirements linked to tests
Accessibility and security test reports
Records of design decisions
Test results for every AI feature
Draft DPIA and transparency record
Runbooks and operating notes
A record of AI and human contributions