Inspiration
Imagine a project manager must urgently deliver a Salesforce application, but several key developers are on vacation. The work still needs to be planned, built, tested, reviewed, approved, and deployed.
NexusDev AI keeps the project moving. Five specialized AI agents work together to manage the software delivery lifecycle, while people approve critical decisions before changes reach staging or production.
What it does
NexusDev AI turns a Jira Epic into a tested, security-reviewed, and deployable Salesforce solution.
Its five agents handle different parts of the workflow:
Decomposer Agent
Breaks the Epic into user stories, technical requirements, acceptance criteria, and development tasks.Developer Agent
Creates Apex and Lightning Web Component code, runs tests, analyzes failures, and attempts corrections.Security Agent
Checks for issues such as SOQL injection, exposed credentials, unsafe permissions, and sensitive-data leaks.GitOps Agent
Creates a feature branch, commits verified Salesforce metadata, and prepares the pull request.Depoyment Agent
Coordinates deployments across the QA sandbox and production organization.
Human approval is required at three important stages:
- Scope and requirements
- Code and security review
- QA sandbox/ Production release
Approvers can approve, reject, or request changes. The agents then update the solution and resubmit it for review.
How we built it
NexusDev AI is built on Google’s AI and Cloud ecosystem:
- Google ADK and Gemini coordinate agent reasoning and tool execution.
- Gemini 3.6 Flash supports planning, GitOps, and release coordination.
- Gemini 3.5 Pro handles complex Apex and LWC development and test-failure analysis.
- Gemma 2 and Model Armor support security and governance checks.
- Cloud Run hosts the containerized services.
- Firestore preserves workflow state while waiting for human approval.
- Pub/Sub and Cloud Scheduler manage events and scheduled audits.
- Google Cloud Observability tracks logs, performance, and execution traces.
The platform connects with Jira, Git, and Salesforce CLI through reusable tools and domain-specific Agent Skills.
Challenges we ran into
Waiting for human approval
Approvals may take minutes or hours, while Cloud Run services are short-lived. We use Firestore to save the workflow state and restore it when an approver responds.
Preventing endless correction loops
An automated fix can sometimes solve one test while breaking another. The Developer Agent receives exact compiler errors and stack traces, with a maximum of three correction attempts.
Reducing cost and latency
Five agents with many tools can consume unnecessary tokens. Our optimizer loads only the tools needed for the current step, reducing overhead and improving response time.
Securing Salesforce authentication
Salesforce CLI runs inside Cloud Run using secure, non-interactive authentication designed for isolated deployment environments.
Accomplishments that we're proud of
NexusDev AI demonstrates an end-to-end workflow that can:
- Convert a Jira Epic into development-ready tasks
- Generate Apex and LWC code
- Run tests and correct common failures
- Perform security and governance checks
- Create a Git feature branch and commit the changes
- Deploy through QA and production
- Pause for human approval at critical stages
The result is faster delivery without giving up human control.
What we learned
Focused Agent Skills work better than oversized prompts.
Precise compiler errors make automated code correction more effective.
Persistent state is essential when workflows include human approvals.
Automation works best when each agent has a clear responsibility.
What's next for NexusDev AI — Autonomous Enterprise workflow Orchestrator
We plan to expand NexusDev AI with:
- Support for ServiceNow, SAP, Kubernetes, and AWS
- Parallel workers for integration and regression testing
- Predictive deployment-risk and dependency analysis
- Approval experiences inside Slack and Microsoft Teams
Log in or sign up for Devpost to join the conversation.