Next Planned Releases
The features below represent the most requested and highest-impact capabilities identified through engineering analysis, user feedback, and SDLC workflow research. Each release is grounded in documented design specifications or research findings in the CPU-Agents-for-SDLC repository.
Automated Test Generation
End-to-end automated test generation from GUI object maps and database schemas, eliminating manual Playwright script authoring.
- GUI Object Mapping Service — DOM acquisition via Playwright, 5-tier selector strategy (ID → data-testid → ARIA → CSS → XPath)
- AI-powered element classification using local Granite 4 / Phi-3 models
- Database Discovery Service — multi-RDBMS schema introspection (SQL Server, PostgreSQL, MySQL, Oracle)
- Page Object Model builder with auto-generated TypeScript fixtures
- Data setup script generator for pre-condition seeding
- DBA Work Item Orchestrator — approval workflow for schema-change test cases
Test authoring accounts for 40+ hours per feature. Automated generation targets a 70% reduction to ~12 hours, directly addressing the highest-cost SDLC activity.
Distributed Agent Execution
Hub-node architecture enabling parallel test execution across multiple workstations, eliminating single-machine bottlenecks.
- Hub controller with work queue and node health monitoring
- Worker node agents for Windows and Linux (Podman rootless)
- Selenium Grid 4-compatible node registration protocol
- FFmpeg-based screen recording for test evidence capture
- Low-latency streaming of test execution video to hub
- Offline queue with SQLite caching and conflict resolution
- OpenTelemetry distributed tracing across all nodes
Sequential execution on a single machine is the primary throughput constraint. Distributed execution across 4–8 nodes multiplies test capacity without additional licensing costs.
Mobile Micro-Agent
Lightweight on-device agent for iOS and Android enabling field engineers to review requirements, capture accessibility issues, and queue work items from mobile.
- Lightweight SLM engine optimized for Apple Neural Engine and Android TPU
- Voice-to-documentation — speech transcription to structured work items
- Camera-based accessibility scanner using on-device vision models
- Quick test case review and approval workflow
- Offline work queue with background sync to Azure DevOps
- Model quantization (INT4/INT8) for sub-2 GB memory footprint
- Context window management for constrained memory environments
Field engineers and QA leads need lightweight access to the agent system without requiring a full workstation. Mobile coverage extends autonomous SDLC to on-site and remote scenarios.
Security Hardening & Supply Chain Controls
Production-grade security controls for the agent runtime, dependency supply chain, and Azure DevOps integration credentials.
- Podman rootless container isolation for all agent processes
- Custom seccomp profile dropping 200+ unnecessary syscalls
- CycloneDX SBOM generation on every build with CVE annotation
- Automated dependency pinning with SHA-256 hash verification
- 5-stage gatekeeping: automated scan → binary detection → manual review → staging canary → production
- PAT rotation automation with Azure Key Vault integration
- Audit log with tamper-evident signatures for compliance (SOC 2, ISO 27001)
Supply chain attacks (xz backdoor, SolarWinds) demonstrate that agent runtimes with privileged DevOps access are high-value targets. Security hardening is a prerequisite for enterprise adoption.
AI Model Upgrade — Granite 4 & Phi-4
Upgrade the local AI decision module to Granite 4 and Phi-4 for improved code analysis accuracy, lower latency, and expanded context windows.
- Granite 4 (8B, INT4 quantized) — 128K context for full codebase analysis
- Phi-4 Mini — sub-500ms inference on Intel Core Ultra for real-time code review
- Llama 3.3 70B via vLLM for batch test generation tasks
- Model router — automatic selection based on task type and hardware profile
- Structured JSON output mode for deterministic acceptance criteria evaluation
- Prompt template library for SDLC-specific tasks (code review, test gen, doc update)
Granite 4 and Phi-4 deliver 2–3× accuracy improvements on code-related benchmarks over Phi-3, enabling higher-confidence autonomous decisions without human review.
Multi-Tenant SaaS Mode
Optional cloud-hosted deployment mode allowing multiple teams to share agent infrastructure with full tenant isolation.
- Tenant-scoped Azure DevOps credential vaults
- Per-tenant agent pools with resource quotas
- Shared model serving layer (vLLM multi-tenant) with request isolation
- Usage metering and cost allocation per team
- Self-service onboarding portal — enterprise network access, no external OAuth dependency
- Cross-tenant analytics dashboard for engineering leadership
Organizations with multiple development teams benefit from shared infrastructure while maintaining strict data isolation. SaaS mode reduces per-team setup from days to minutes.
Shape the Roadmap
Feature priorities are driven by real SDLC pain points. If you have a use case not covered above, open an issue or discussion in the GitHub repository. All design documents are public and contributions are welcome.