Fifteen small compute modules working in parallel, their light flowing through a glass approval gate into a screen of code

Multi-Agent Delivery System

Software & SaaS
Contact Us
Software & SaaS

Multi-Agent Delivery System

RemSoft.Dev built a multi-agent software development system in which fifteen AI agents work in parallel to transform detailed specifications into production-ready software. Orchestrated on .NET 10 and Aspire, the platform turns Azure DevOps, Jira or free-text work items into specifications, code, tests and pull requests. A dependency-aware scheduler routes complex reasoning to Claude, Codex or Gemini and implementation tasks to local models; Rust components with AMD drivers and ROCm run those models through WSL for maximum local performance. GraphRAG and code-RAG supply context, human approval gates keep engineers in control, and token and cost tracking make every run accountable. A dedicated test framework validates agent flows, model routing and system stability. The system was tested successfully on two real commercial projects, significantly reducing development time - and it is the delivery engine behind RemSoft.Dev's own client work.

AI agentsDeveloper toolingOrchestrationLLM

Key Challenges

Client: RemSoft.Dev platform, used on client projects

Agents: 15 - Claude Code, Codex, Gemini, local LLMs

Stack: .NET 10 / Aspire, Rust, ROCm/WSL, GraphRAG

Validation: Dedicated agent-flow test framework

The project faced several key challenges:

  • Orchestrating unlike models: Cloud agents and small local models have different strengths and failure modes - the scheduler had to route work by task type and dependency, not just availability.
  • Keeping humans in charge: Approval gates and review checkpoints are part of the architecture: nothing merges because an agent said so.
  • Local performance: Getting real throughput from local models on AMD hardware required Rust components and a ROCm/WSL execution path.
  • Proving it on real work: The platform was validated on two commercial projects, not benchmarks.

Results

  • 15 orchestrated AI agents
  • Work item → spec → code → tests → PR
  • Cloud/local model routing
  • Human approval gates + cost tracking
  • Proven on 2 commercial projects

15 agents

collaborating on planning, coding, testing and review.

2 projects

commercial deliveries validated the system.

Significant

reduction in development time on real work.

Related case studies

Want to Build Something Similar?

Tell us about your project - we will propose an approach, a timeline and a first estimate.

Contact Us
An unhandled error has occurred. Reload 🗙