For CIOs and CTOs
Increase delivery capacity without introducing an ungoverned path around architecture, identity or release management.
Secure AI Software Engineering Agents
AI coding agents inside enterprise controls
Give development teams agentic engineering capability without giving an autonomous process unrestricted access to source code, secrets, networks, CI/CD or production.
Engineering capability
A delivery agent, not an unsupervised developer
The agent can understand a repository, plan a change, edit code, run tools and prepare a pull request. It remains inside the same engineering and security controls expected of a production team.
For engineering leaders
Accelerate backlog, migration, testing and maintenance while preserving review standards and team ownership.
For security teams
Define model, repository, network, secret and tool boundaries with traceable actions and enforceable controls.
For government delivery teams
Use agent assistance in customer-controlled environments with explicit data handling and human accountability.
Use cases
Secure assistance across the software lifecycle
The operating model can support advisory, drafting and approval-required work across application, data and platform engineering.
Repository understanding
Map unfamiliar codebases, dependencies, ownership boundaries and runtime behaviour.
Code generation and refactoring
Prepare scoped implementation changes that follow repository conventions and architecture.
Tests and documentation
Create focused tests, technical documentation and migration notes alongside code changes.
Review and security analysis
Inspect changes for defects, insecure patterns, dependency risk and missing controls.
Infrastructure and CI/CD
Generate and review Bicep, Terraform, pipelines and deployment configuration within policy.
Issue resolution
Reproduce failures, inspect logs, diagnose root causes and prepare reviewable fixes.
Migration and modernisation
Analyse legacy applications, plan incremental change and automate repetitive conversion work.
Pull request delivery
Work on an isolated branch, run checks and open a traceable pull request for human review.
Controlled delivery flow
From issue to pull request with authority still in human hands
The agent works in an isolated environment, on a scoped branch, and produces the same reviewable artefacts expected from a disciplined engineering workflow.
Secure software engineering agent workflow
DW Data reference patternDeveloper
AI engineering agent
Isolated environment
Scoped branch
Tests
Security scan
Pull request
Human reviewer
CI/CD
Deployment
Agent plans and prepares
The model can reason over the issue, repository and test results, then propose a bounded change.
Controls inspect and constrain
Environment policy, schemas, scanners and permissions determine what can actually run.
People retain authority
A human reviews the code and evidence. Existing branch and deployment approvals remain authoritative.
Security architecture
A bounded execution environment for every task
Controls are applied at the identity, repository, compute, network, secret, model and release boundaries.
Repository-scoped access
The agent receives access only to approved repositories, branches and issue context. Cross-repository access is explicit.
Isolated execution
Commands run in an ephemeral, monitored environment with defined compute, filesystem and persistence boundaries.
Controlled network egress
Outbound access is denied or allow-listed. Package registries, documentation and tools are approved by environment policy.
Brokered secrets
Secrets are not placed in prompts. Short-lived credentials are injected only for approved operations at runtime.
Mandatory security checks
Static analysis, secret scanning, dependency analysis, tests and policy checks run before a pull request can progress.
Human merge and release authority
The agent cannot approve or merge its own pull request and cannot bypass existing production release controls.
Explicit deny policy
What the coding agent cannot do
These are enforceable platform controls, not behavioural requests hidden inside a prompt.
- Merge or approve its own pull request
- Deploy directly to production
- Retrieve arbitrary secrets or credentials
- Disable branch protection or security scanning
- Expand its own repository or network scope
- Send source code to an unapproved model
- Suppress failed checks without an authorised decision
- Modify audit records or its own control policy
Traceability
Evidence from intent to deployment
Every material step can be joined into a single delivery trace for engineering, security and audit review.
Identity
Developer, agent and execution-environment identity for each run
Inputs
Issue, repository, selected files, instructions and model route
Actions
Commands, file changes, tool calls, network destinations and results
Quality
Test results, coverage changes, lint, build and evaluation evidence
Security
Secret scan, SAST, dependency and infrastructure policy results
Decision
Pull request, human review, approval, merge and deployment record
Model and data policy
Source code follows the same sovereignty decision as other sensitive data
Repository content is classified before model access. The DW Model Gateway restricts sensitive code to approved deployments and allows newer external coding models only where customer policy explicitly permits them.
High-control repository
Approved regional model, restricted egress, private execution and no external model route.
Internal application code
Customer-approved model catalogue, repository-scoped access and controlled package sources.
Public open-source context
Frontier coding model may be permitted with explicit policy and provider terms.
Mixed classification task
Minimise or exclude sensitive files, split the task and route each part to an approved boundary.
Implementation model
A practical path to a production coding agent
Start with a constrained workflow and earn broader capability through measured evidence.
- 01
Select the workflow
Choose a bounded repository task with clear success, risk and review criteria.
- 02
Build the sandbox
Configure identity, repository scope, compute, tools, egress and secret brokering.
- 03
Evaluate on real work
Benchmark quality, security, cost and review effort against representative issues.
- 04
Operate and expand
Monitor production use and add capabilities only when control evidence supports it.
From concept to controlled production
Add coding agents without weakening engineering control
DW Data can design the model boundary, isolated runtime, repository permissions, security checks, pull request workflow and AgentOps needed for a secure production service.